A666sxr Cybercat commited on
Commit
e0b0e2f
0 Parent(s):

Duplicate from Cybercat/Genshin_MB_VITS_TTS

Browse files

Co-authored-by: CyberCat <Cybercat@users.noreply.huggingface.co>

.gitattributes ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tflite filter=lfs diff=lfs merge=lfs -text
29
+ *.tgz filter=lfs diff=lfs merge=lfs -text
30
+ *.wasm filter=lfs diff=lfs merge=lfs -text
31
+ *.xz filter=lfs diff=lfs merge=lfs -text
32
+ *.zip filter=lfs diff=lfs merge=lfs -text
33
+ *.zst filter=lfs diff=lfs merge=lfs -text
34
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6
+
7
+ 1. Definitions.
8
+
9
+ "License" shall mean the terms and conditions for use, reproduction,
10
+ and distribution as defined by Sections 1 through 9 of this document.
11
+
12
+ "Licensor" shall mean the copyright owner or entity authorized by
13
+ the copyright owner that is granting the License.
14
+
15
+ "Legal Entity" shall mean the union of the acting entity and all
16
+ other entities that control, are controlled by, or are under common
17
+ control with that entity. For the purposes of this definition,
18
+ "control" means (i) the power, direct or indirect, to cause the
19
+ direction or management of such entity, whether by contract or
20
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21
+ outstanding shares, or (iii) beneficial ownership of such entity.
22
+
23
+ "You" (or "Your") shall mean an individual or Legal Entity
24
+ exercising permissions granted by this License.
25
+
26
+ "Source" form shall mean the preferred form for making modifications,
27
+ including but not limited to software source code, documentation
28
+ source, and configuration files.
29
+
30
+ "Object" form shall mean any form resulting from mechanical
31
+ transformation or translation of a Source form, including but
32
+ not limited to compiled object code, generated documentation,
33
+ and conversions to other media types.
34
+
35
+ "Work" shall mean the work of authorship, whether in Source or
36
+ Object form, made available under the License, as indicated by a
37
+ copyright notice that is included in or attached to the work
38
+ (an example is provided in the Appendix below).
39
+
40
+ "Derivative Works" shall mean any work, whether in Source or Object
41
+ form, that is based on (or derived from) the Work and for which the
42
+ editorial revisions, annotations, elaborations, or other modifications
43
+ represent, as a whole, an original work of authorship. For the purposes
44
+ of this License, Derivative Works shall not include works that remain
45
+ separable from, or merely link (or bind by name) to the interfaces of,
46
+ the Work and Derivative Works thereof.
47
+
48
+ "Contribution" shall mean any work of authorship, including
49
+ the original version of the Work and any modifications or additions
50
+ to that Work or Derivative Works thereof, that is intentionally
51
+ submitted to Licensor for inclusion in the Work by the copyright owner
52
+ or by an individual or Legal Entity authorized to submit on behalf of
53
+ the copyright owner. For the purposes of this definition, "submitted"
54
+ means any form of electronic, verbal, or written communication sent
55
+ to the Licensor or its representatives, including but not limited to
56
+ communication on electronic mailing lists, source code control systems,
57
+ and issue tracking systems that are managed by, or on behalf of, the
58
+ Licensor for the purpose of discussing and improving the Work, but
59
+ excluding communication that is conspicuously marked or otherwise
60
+ designated in writing by the copyright owner as "Not a Contribution."
61
+
62
+ "Contributor" shall mean Licensor and any individual or Legal Entity
63
+ on behalf of whom a Contribution has been received by Licensor and
64
+ subsequently incorporated within the Work.
65
+
66
+ 2. Grant of Copyright License. Subject to the terms and conditions of
67
+ this License, each Contributor hereby grants to You a perpetual,
68
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69
+ copyright license to reproduce, prepare Derivative Works of,
70
+ publicly display, publicly perform, sublicense, and distribute the
71
+ Work and such Derivative Works in Source or Object form.
72
+
73
+ 3. Grant of Patent License. Subject to the terms and conditions of
74
+ this License, each Contributor hereby grants to You a perpetual,
75
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76
+ (except as stated in this section) patent license to make, have made,
77
+ use, offer to sell, sell, import, and otherwise transfer the Work,
78
+ where such license applies only to those patent claims licensable
79
+ by such Contributor that are necessarily infringed by their
80
+ Contribution(s) alone or by combination of their Contribution(s)
81
+ with the Work to which such Contribution(s) was submitted. If You
82
+ institute patent litigation against any entity (including a
83
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84
+ or a Contribution incorporated within the Work constitutes direct
85
+ or contributory patent infringement, then any patent licenses
86
+ granted to You under this License for that Work shall terminate
87
+ as of the date such litigation is filed.
88
+
89
+ 4. Redistribution. You may reproduce and distribute copies of the
90
+ Work or Derivative Works thereof in any medium, with or without
91
+ modifications, and in Source or Object form, provided that You
92
+ meet the following conditions:
93
+
94
+ (a) You must give any other recipients of the Work or
95
+ Derivative Works a copy of this License; and
96
+
97
+ (b) You must cause any modified files to carry prominent notices
98
+ stating that You changed the files; and
99
+
100
+ (c) You must retain, in the Source form of any Derivative Works
101
+ that You distribute, all copyright, patent, trademark, and
102
+ attribution notices from the Source form of the Work,
103
+ excluding those notices that do not pertain to any part of
104
+ the Derivative Works; and
105
+
106
+ (d) If the Work includes a "NOTICE" text file as part of its
107
+ distribution, then any Derivative Works that You distribute must
108
+ include a readable copy of the attribution notices contained
109
+ within such NOTICE file, excluding those notices that do not
110
+ pertain to any part of the Derivative Works, in at least one
111
+ of the following places: within a NOTICE text file distributed
112
+ as part of the Derivative Works; within the Source form or
113
+ documentation, if provided along with the Derivative Works; or,
114
+ within a display generated by the Derivative Works, if and
115
+ wherever such third-party notices normally appear. The contents
116
+ of the NOTICE file are for informational purposes only and
117
+ do not modify the License. You may add Your own attribution
118
+ notices within Derivative Works that You distribute, alongside
119
+ or as an addendum to the NOTICE text from the Work, provided
120
+ that such additional attribution notices cannot be construed
121
+ as modifying the License.
122
+
123
+ You may add Your own copyright statement to Your modifications and
124
+ may provide additional or different license terms and conditions
125
+ for use, reproduction, or distribution of Your modifications, or
126
+ for any such Derivative Works as a whole, provided Your use,
127
+ reproduction, and distribution of the Work otherwise complies with
128
+ the conditions stated in this License.
129
+
130
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131
+ any Contribution intentionally submitted for inclusion in the Work
132
+ by You to the Licensor shall be under the terms and conditions of
133
+ this License, without any additional terms or conditions.
134
+ Notwithstanding the above, nothing herein shall supersede or modify
135
+ the terms of any separate license agreement you may have executed
136
+ with Licensor regarding such Contributions.
137
+
138
+ 6. Trademarks. This License does not grant permission to use the trade
139
+ names, trademarks, service marks, or product names of the Licensor,
140
+ except as required for reasonable and customary use in describing the
141
+ origin of the Work and reproducing the content of the NOTICE file.
142
+
143
+ 7. Disclaimer of Warranty. Unless required by applicable law or
144
+ agreed to in writing, Licensor provides the Work (and each
145
+ Contributor provides its Contributions) on an "AS IS" BASIS,
146
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147
+ implied, including, without limitation, any warranties or conditions
148
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149
+ PARTICULAR PURPOSE. You are solely responsible for determining the
150
+ appropriateness of using or redistributing the Work and assume any
151
+ risks associated with Your exercise of permissions under this License.
152
+
153
+ 8. Limitation of Liability. In no event and under no legal theory,
154
+ whether in tort (including negligence), contract, or otherwise,
155
+ unless required by applicable law (such as deliberate and grossly
156
+ negligent acts) or agreed to in writing, shall any Contributor be
157
+ liable to You for damages, including any direct, indirect, special,
158
+ incidental, or consequential damages of any character arising as a
159
+ result of this License or out of the use or inability to use the
160
+ Work (including but not limited to damages for loss of goodwill,
161
+ work stoppage, computer failure or malfunction, or any and all
162
+ other commercial damages or losses), even if such Contributor
163
+ has been advised of the possibility of such damages.
164
+
165
+ 9. Accepting Warranty or Additional Liability. While redistributing
166
+ the Work or Derivative Works thereof, You may choose to offer,
167
+ and charge a fee for, acceptance of support, warranty, indemnity,
168
+ or other liability obligations and/or rights consistent with this
169
+ License. However, in accepting such obligations, You may act only
170
+ on Your own behalf and on Your sole responsibility, not on behalf
171
+ of any other Contributor, and only if You agree to indemnify,
172
+ defend, and hold each Contributor harmless for any liability
173
+ incurred by, or claims asserted against, such Contributor by reason
174
+ of your accepting any such warranty or additional liability.
175
+
176
+ END OF TERMS AND CONDITIONS
177
+
178
+ APPENDIX: How to apply the Apache License to your work.
179
+
180
+ To apply the Apache License to your work, attach the following
181
+ boilerplate notice, with the fields enclosed by brackets "[]"
182
+ replaced with your own identifying information. (Don't include
183
+ the brackets!) The text should be enclosed in the appropriate
184
+ comment syntax for the file format. We also recommend that a
185
+ file or class name and description of purpose be included on the
186
+ same "printed page" as the copyright notice for easier
187
+ identification within third-party archives.
188
+
189
+ Copyright [yyyy] [name of copyright owner]
190
+
191
+ Licensed under the Apache License, Version 2.0 (the "License");
192
+ you may not use this file except in compliance with the License.
193
+ You may obtain a copy of the License at
194
+
195
+ http://www.apache.org/licenses/LICENSE-2.0
196
+
197
+ Unless required by applicable law or agreed to in writing, software
198
+ distributed under the License is distributed on an "AS IS" BASIS,
199
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200
+ See the License for the specific language governing permissions and
201
+ limitations under the License.
README.md ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: Genshin TTS
3
+ emoji: 🔥
4
+ colorFrom: pink
5
+ colorTo: red
6
+ sdk: gradio
7
+ sdk_version: 3.12.0
8
+ app_file: app.py
9
+ pinned: false
10
+ duplicated_from: Cybercat/Genshin_MB_VITS_TTS
11
+ ---
12
+
13
+ Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
app.py ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import os
3
+ os.system('cd monotonic_align && python setup.py build_ext --inplace && cd ..')
4
+
5
+ import time
6
+ import json
7
+ import math
8
+ import torch
9
+ from torch import nn
10
+ from torch.nn import functional as F
11
+ from torch.utils.data import DataLoader
12
+ import re
13
+ import langid
14
+ import jieba
15
+ import commons
16
+ import utils
17
+ from data_utils import TextAudioLoader, TextAudioCollate, TextAudioSpeakerLoader, TextAudioSpeakerCollate
18
+ from models import SynthesizerTrn
19
+ from text.symbols import symbols
20
+ from text import text_to_sequence, cleaned_text_to_sequence
21
+ from text.cleaners import japanese_cleaners
22
+ from scipy.io.wavfile import write
23
+
24
+ def getMixText(text):
25
+ langid.set_languages(['zh','en'])
26
+ seg_list = jieba.cut(text, cut_all=False)
27
+ clean_list=[]
28
+ for seg in seg_list:
29
+ langtext='[ZH]'
30
+ if(len(seg)>0):
31
+ lang=langid.classify(seg)[0]
32
+ if lang == 'en':
33
+ langtext='[EN]'
34
+ elif lang=='zh':
35
+ langtext='[ZH]'
36
+ clean_list.append(langtext+seg+langtext)
37
+ return ''.join(clean_list)
38
+
39
+ def get_text(text, hps):
40
+ text_norm = text_to_sequence(text, hps.data.text_cleaners)
41
+ if hps.data.add_blank:
42
+ text_norm = commons.intersperse(text_norm, 0)
43
+ text_norm = torch.LongTensor(text_norm)
44
+ return text_norm
45
+
46
+ hps_ms = utils.get_hparams_from_file("save_model/config.json")
47
+ hps = utils.get_hparams_from_file("save_model/config.json")
48
+ net_g_ms = SynthesizerTrn(
49
+ len(symbols),
50
+ hps_ms.data.filter_length // 2 + 1,
51
+ hps_ms.train.segment_size // hps.data.hop_length,
52
+ n_speakers=hps_ms.data.n_speakers,
53
+ **hps_ms.model)
54
+
55
+ npclists=[]
56
+ with open("save_model/npclists.txt",'r') as r:
57
+ for npc in r.readlines():
58
+ npclists.append(npc.split('|')[-1])
59
+ print(npc)
60
+ r.close
61
+
62
+ def tts(spkid, text):
63
+ if(len(re.findall(r'\[ZH\].*?\[ZH\]', text))==0 and len(re.findall(r'\[EN\].*?\[EN\]', text))==0):
64
+ text=getMixText(text)
65
+ sid = torch.LongTensor([spkid]) # speaker identity
66
+ stn_tst = get_text(text, hps_ms)
67
+
68
+ with torch.no_grad():
69
+ x_tst = stn_tst.unsqueeze(0)
70
+ x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
71
+ t1 = time.time()
72
+ audio = net_g_ms.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8, length_scale=1)[0][
73
+ 0, 0].data.float().numpy()
74
+ t2 = time.time()
75
+ return "成功,耗时"+str((t2-t1))+"s", (hps.data.sampling_rate, audio)
76
+
77
+
78
+ _ = utils.load_checkpoint("save_model/model.pth", net_g_ms, None)
79
+
80
+ def clean_text(text):
81
+ return japanese_cleaners(text)
82
+
83
+ app = gr.Blocks()
84
+ with app:
85
+ with gr.Tabs():
86
+ with gr.TabItem("Basic"):
87
+ tts_input1 = gr.TextArea(label="在这输入文字", value="在我的后园,可以看见墙外有两株树,一株是枣树,还有一株也是枣树。 这上面的夜的天空,奇怪而高,我生平没有见过这样的奇怪而高的天空。 他仿佛要离开人间而去,使人们仰面不再看见。 然而现在却非常之蓝,闪闪地眨着几十个星星的眼,冷眼。 他的口角上现出微笑,似乎自以为大有深意,而将繁霜洒在我的园里的野花草上。 我不知道那些花草真叫什么名字,人们叫他们什么名字。 我记得有一种开过极细小的粉红花,现在还开着,但是更极细小了,她在冷的夜气中,瑟缩地做梦,梦见春的到来,梦见秋的到来,梦见瘦的诗人将眼泪擦在她最末的花瓣上,告诉她秋虽然来,冬虽然来,而此后接着还是春,胡蝶乱飞,蜜蜂都唱起春词来了。 她于是一笑,虽然颜色冻得红惨惨地,仍然瑟缩着。 枣树,他们简直落尽了叶子。")
88
+ tts_input2 = gr.Dropdown(label="人物", choices=npclists, type="index", value=npclists[0])
89
+ tts_submit = gr.Button("合成", variant="primary")
90
+ tts_output1 = gr.Textbox(label="信息")
91
+ tts_output2 = gr.Audio(label="结果")
92
+ tts_submit.click(tts, [tts_input2, tts_input1], [tts_output1, tts_output2])
93
+ app.launch()
attentions.py ADDED
@@ -0,0 +1,303 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import numpy as np
4
+ import torch
5
+ from torch import nn
6
+ from torch.nn import functional as F
7
+
8
+ import commons
9
+ import modules
10
+ from modules import LayerNorm
11
+
12
+
13
+ class Encoder(nn.Module):
14
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
15
+ super().__init__()
16
+ self.hidden_channels = hidden_channels
17
+ self.filter_channels = filter_channels
18
+ self.n_heads = n_heads
19
+ self.n_layers = n_layers
20
+ self.kernel_size = kernel_size
21
+ self.p_dropout = p_dropout
22
+ self.window_size = window_size
23
+
24
+ self.drop = nn.Dropout(p_dropout)
25
+ self.attn_layers = nn.ModuleList()
26
+ self.norm_layers_1 = nn.ModuleList()
27
+ self.ffn_layers = nn.ModuleList()
28
+ self.norm_layers_2 = nn.ModuleList()
29
+ for i in range(self.n_layers):
30
+ self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
31
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
32
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
33
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
34
+
35
+ def forward(self, x, x_mask):
36
+ attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
37
+ x = x * x_mask
38
+ for i in range(self.n_layers):
39
+ y = self.attn_layers[i](x, x, attn_mask)
40
+ y = self.drop(y)
41
+ x = self.norm_layers_1[i](x + y)
42
+
43
+ y = self.ffn_layers[i](x, x_mask)
44
+ y = self.drop(y)
45
+ x = self.norm_layers_2[i](x + y)
46
+ x = x * x_mask
47
+ return x
48
+
49
+
50
+ class Decoder(nn.Module):
51
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
52
+ super().__init__()
53
+ self.hidden_channels = hidden_channels
54
+ self.filter_channels = filter_channels
55
+ self.n_heads = n_heads
56
+ self.n_layers = n_layers
57
+ self.kernel_size = kernel_size
58
+ self.p_dropout = p_dropout
59
+ self.proximal_bias = proximal_bias
60
+ self.proximal_init = proximal_init
61
+
62
+ self.drop = nn.Dropout(p_dropout)
63
+ self.self_attn_layers = nn.ModuleList()
64
+ self.norm_layers_0 = nn.ModuleList()
65
+ self.encdec_attn_layers = nn.ModuleList()
66
+ self.norm_layers_1 = nn.ModuleList()
67
+ self.ffn_layers = nn.ModuleList()
68
+ self.norm_layers_2 = nn.ModuleList()
69
+ for i in range(self.n_layers):
70
+ self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
71
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
72
+ self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
73
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
74
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
75
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
76
+
77
+ def forward(self, x, x_mask, h, h_mask):
78
+ """
79
+ x: decoder input
80
+ h: encoder output
81
+ """
82
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
83
+ encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
84
+ x = x * x_mask
85
+ for i in range(self.n_layers):
86
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
87
+ y = self.drop(y)
88
+ x = self.norm_layers_0[i](x + y)
89
+
90
+ y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
91
+ y = self.drop(y)
92
+ x = self.norm_layers_1[i](x + y)
93
+
94
+ y = self.ffn_layers[i](x, x_mask)
95
+ y = self.drop(y)
96
+ x = self.norm_layers_2[i](x + y)
97
+ x = x * x_mask
98
+ return x
99
+
100
+
101
+ class MultiHeadAttention(nn.Module):
102
+ def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
103
+ super().__init__()
104
+ assert channels % n_heads == 0
105
+
106
+ self.channels = channels
107
+ self.out_channels = out_channels
108
+ self.n_heads = n_heads
109
+ self.p_dropout = p_dropout
110
+ self.window_size = window_size
111
+ self.heads_share = heads_share
112
+ self.block_length = block_length
113
+ self.proximal_bias = proximal_bias
114
+ self.proximal_init = proximal_init
115
+ self.attn = None
116
+
117
+ self.k_channels = channels // n_heads
118
+ self.conv_q = nn.Conv1d(channels, channels, 1)
119
+ self.conv_k = nn.Conv1d(channels, channels, 1)
120
+ self.conv_v = nn.Conv1d(channels, channels, 1)
121
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
122
+ self.drop = nn.Dropout(p_dropout)
123
+
124
+ if window_size is not None:
125
+ n_heads_rel = 1 if heads_share else n_heads
126
+ rel_stddev = self.k_channels**-0.5
127
+ self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
128
+ self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
129
+
130
+ nn.init.xavier_uniform_(self.conv_q.weight)
131
+ nn.init.xavier_uniform_(self.conv_k.weight)
132
+ nn.init.xavier_uniform_(self.conv_v.weight)
133
+ if proximal_init:
134
+ with torch.no_grad():
135
+ self.conv_k.weight.copy_(self.conv_q.weight)
136
+ self.conv_k.bias.copy_(self.conv_q.bias)
137
+
138
+ def forward(self, x, c, attn_mask=None):
139
+ q = self.conv_q(x)
140
+ k = self.conv_k(c)
141
+ v = self.conv_v(c)
142
+
143
+ x, self.attn = self.attention(q, k, v, mask=attn_mask)
144
+
145
+ x = self.conv_o(x)
146
+ return x
147
+
148
+ def attention(self, query, key, value, mask=None):
149
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
150
+ b, d, t_s, t_t = (*key.size(), query.size(2))
151
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
152
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
153
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
154
+
155
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
156
+ if self.window_size is not None:
157
+ assert t_s == t_t, "Relative attention is only available for self-attention."
158
+ key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
159
+ rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
160
+ scores_local = self._relative_position_to_absolute_position(rel_logits)
161
+ scores = scores + scores_local
162
+ if self.proximal_bias:
163
+ assert t_s == t_t, "Proximal bias is only available for self-attention."
164
+ scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
165
+ if mask is not None:
166
+ scores = scores.masked_fill(mask == 0, -1e4)
167
+ if self.block_length is not None:
168
+ assert t_s == t_t, "Local attention is only available for self-attention."
169
+ block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
170
+ scores = scores.masked_fill(block_mask == 0, -1e4)
171
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
172
+ p_attn = self.drop(p_attn)
173
+ output = torch.matmul(p_attn, value)
174
+ if self.window_size is not None:
175
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
176
+ value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
177
+ output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
178
+ output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
179
+ return output, p_attn
180
+
181
+ def _matmul_with_relative_values(self, x, y):
182
+ """
183
+ x: [b, h, l, m]
184
+ y: [h or 1, m, d]
185
+ ret: [b, h, l, d]
186
+ """
187
+ ret = torch.matmul(x, y.unsqueeze(0))
188
+ return ret
189
+
190
+ def _matmul_with_relative_keys(self, x, y):
191
+ """
192
+ x: [b, h, l, d]
193
+ y: [h or 1, m, d]
194
+ ret: [b, h, l, m]
195
+ """
196
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
197
+ return ret
198
+
199
+ def _get_relative_embeddings(self, relative_embeddings, length):
200
+ max_relative_position = 2 * self.window_size + 1
201
+ # Pad first before slice to avoid using cond ops.
202
+ pad_length = max(length - (self.window_size + 1), 0)
203
+ slice_start_position = max((self.window_size + 1) - length, 0)
204
+ slice_end_position = slice_start_position + 2 * length - 1
205
+ if pad_length > 0:
206
+ padded_relative_embeddings = F.pad(
207
+ relative_embeddings,
208
+ commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
209
+ else:
210
+ padded_relative_embeddings = relative_embeddings
211
+ used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
212
+ return used_relative_embeddings
213
+
214
+ def _relative_position_to_absolute_position(self, x):
215
+ """
216
+ x: [b, h, l, 2*l-1]
217
+ ret: [b, h, l, l]
218
+ """
219
+ batch, heads, length, _ = x.size()
220
+ # Concat columns of pad to shift from relative to absolute indexing.
221
+ x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
222
+
223
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
224
+ x_flat = x.view([batch, heads, length * 2 * length])
225
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
226
+
227
+ # Reshape and slice out the padded elements.
228
+ x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
229
+ return x_final
230
+
231
+ def _absolute_position_to_relative_position(self, x):
232
+ """
233
+ x: [b, h, l, l]
234
+ ret: [b, h, l, 2*l-1]
235
+ """
236
+ batch, heads, length, _ = x.size()
237
+ # padd along column
238
+ x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
239
+ x_flat = x.view([batch, heads, length**2 + length*(length -1)])
240
+ # add 0's in the beginning that will skew the elements after reshape
241
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
242
+ x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
243
+ return x_final
244
+
245
+ def _attention_bias_proximal(self, length):
246
+ """Bias for self-attention to encourage attention to close positions.
247
+ Args:
248
+ length: an integer scalar.
249
+ Returns:
250
+ a Tensor with shape [1, 1, length, length]
251
+ """
252
+ r = torch.arange(length, dtype=torch.float32)
253
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
254
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
255
+
256
+
257
+ class FFN(nn.Module):
258
+ def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
259
+ super().__init__()
260
+ self.in_channels = in_channels
261
+ self.out_channels = out_channels
262
+ self.filter_channels = filter_channels
263
+ self.kernel_size = kernel_size
264
+ self.p_dropout = p_dropout
265
+ self.activation = activation
266
+ self.causal = causal
267
+
268
+ if causal:
269
+ self.padding = self._causal_padding
270
+ else:
271
+ self.padding = self._same_padding
272
+
273
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
274
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
275
+ self.drop = nn.Dropout(p_dropout)
276
+
277
+ def forward(self, x, x_mask):
278
+ x = self.conv_1(self.padding(x * x_mask))
279
+ if self.activation == "gelu":
280
+ x = x * torch.sigmoid(1.702 * x)
281
+ else:
282
+ x = torch.relu(x)
283
+ x = self.drop(x)
284
+ x = self.conv_2(self.padding(x * x_mask))
285
+ return x * x_mask
286
+
287
+ def _causal_padding(self, x):
288
+ if self.kernel_size == 1:
289
+ return x
290
+ pad_l = self.kernel_size - 1
291
+ pad_r = 0
292
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
293
+ x = F.pad(x, commons.convert_pad_shape(padding))
294
+ return x
295
+
296
+ def _same_padding(self, x):
297
+ if self.kernel_size == 1:
298
+ return x
299
+ pad_l = (self.kernel_size - 1) // 2
300
+ pad_r = self.kernel_size // 2
301
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
302
+ x = F.pad(x, commons.convert_pad_shape(padding))
303
+ return x
commons.py ADDED
@@ -0,0 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import numpy as np
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+
8
+ def init_weights(m, mean=0.0, std=0.01):
9
+ classname = m.__class__.__name__
10
+ if classname.find("Conv") != -1:
11
+ m.weight.data.normal_(mean, std)
12
+
13
+
14
+ def get_padding(kernel_size, dilation=1):
15
+ return int((kernel_size*dilation - dilation)/2)
16
+
17
+
18
+ def convert_pad_shape(pad_shape):
19
+ l = pad_shape[::-1]
20
+ pad_shape = [item for sublist in l for item in sublist]
21
+ return pad_shape
22
+
23
+
24
+ def intersperse(lst, item):
25
+ result = [item] * (len(lst) * 2 + 1)
26
+ result[1::2] = lst
27
+ return result
28
+
29
+
30
+ def kl_divergence(m_p, logs_p, m_q, logs_q):
31
+ """KL(P||Q)"""
32
+ kl = (logs_q - logs_p) - 0.5
33
+ kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
34
+ return kl
35
+
36
+
37
+ def rand_gumbel(shape):
38
+ """Sample from the Gumbel distribution, protect from overflows."""
39
+ uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
40
+ return -torch.log(-torch.log(uniform_samples))
41
+
42
+
43
+ def rand_gumbel_like(x):
44
+ g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
45
+ return g
46
+
47
+
48
+ def slice_segments(x, ids_str, segment_size=4):
49
+ ret = torch.zeros_like(x[:, :, :segment_size])
50
+ for i in range(x.size(0)):
51
+ idx_str = ids_str[i]
52
+ idx_end = idx_str + segment_size
53
+ ret[i] = x[i, :, idx_str:idx_end]
54
+ return ret
55
+
56
+
57
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
58
+ b, d, t = x.size()
59
+ if x_lengths is None:
60
+ x_lengths = t
61
+ ids_str_max = x_lengths - segment_size + 1
62
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
63
+ ret = slice_segments(x, ids_str, segment_size)
64
+ return ret, ids_str
65
+
66
+
67
+ def get_timing_signal_1d(
68
+ length, channels, min_timescale=1.0, max_timescale=1.0e4):
69
+ position = torch.arange(length, dtype=torch.float)
70
+ num_timescales = channels // 2
71
+ log_timescale_increment = (
72
+ math.log(float(max_timescale) / float(min_timescale)) /
73
+ (num_timescales - 1))
74
+ inv_timescales = min_timescale * torch.exp(
75
+ torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
76
+ scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
77
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
78
+ signal = F.pad(signal, [0, 0, 0, channels % 2])
79
+ signal = signal.view(1, channels, length)
80
+ return signal
81
+
82
+
83
+ def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
84
+ b, channels, length = x.size()
85
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
86
+ return x + signal.to(dtype=x.dtype, device=x.device)
87
+
88
+
89
+ def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
90
+ b, channels, length = x.size()
91
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
92
+ return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
93
+
94
+
95
+ def subsequent_mask(length):
96
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
97
+ return mask
98
+
99
+
100
+ @torch.jit.script
101
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
102
+ n_channels_int = n_channels[0]
103
+ in_act = input_a + input_b
104
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
105
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
106
+ acts = t_act * s_act
107
+ return acts
108
+
109
+
110
+ def convert_pad_shape(pad_shape):
111
+ l = pad_shape[::-1]
112
+ pad_shape = [item for sublist in l for item in sublist]
113
+ return pad_shape
114
+
115
+
116
+ def shift_1d(x):
117
+ x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
118
+ return x
119
+
120
+
121
+ def sequence_mask(length, max_length=None):
122
+ if max_length is None:
123
+ max_length = length.max()
124
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
125
+ return x.unsqueeze(0) < length.unsqueeze(1)
126
+
127
+
128
+ def generate_path(duration, mask):
129
+ """
130
+ duration: [b, 1, t_x]
131
+ mask: [b, 1, t_y, t_x]
132
+ """
133
+ device = duration.device
134
+
135
+ b, _, t_y, t_x = mask.shape
136
+ cum_duration = torch.cumsum(duration, -1)
137
+
138
+ cum_duration_flat = cum_duration.view(b * t_x)
139
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
140
+ path = path.view(b, t_x, t_y)
141
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
142
+ path = path.unsqueeze(1).transpose(2,3) * mask
143
+ return path
144
+
145
+
146
+ def clip_grad_value_(parameters, clip_value, norm_type=2):
147
+ if isinstance(parameters, torch.Tensor):
148
+ parameters = [parameters]
149
+ parameters = list(filter(lambda p: p.grad is not None, parameters))
150
+ norm_type = float(norm_type)
151
+ if clip_value is not None:
152
+ clip_value = float(clip_value)
153
+
154
+ total_norm = 0
155
+ for p in parameters:
156
+ param_norm = p.grad.data.norm(norm_type)
157
+ total_norm += param_norm.item() ** norm_type
158
+ if clip_value is not None:
159
+ p.grad.data.clamp_(min=-clip_value, max=clip_value)
160
+ total_norm = total_norm ** (1. / norm_type)
161
+ return total_norm
data_utils.py ADDED
@@ -0,0 +1,392 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import time
2
+ import os
3
+ import random
4
+ import numpy as np
5
+ import torch
6
+ import torch.utils.data
7
+
8
+ import commons
9
+ from mel_processing import spectrogram_torch
10
+ from utils import load_wav_to_torch, load_filepaths_and_text
11
+ from text import text_to_sequence, cleaned_text_to_sequence
12
+
13
+
14
+ class TextAudioLoader(torch.utils.data.Dataset):
15
+ """
16
+ 1) loads audio, text pairs
17
+ 2) normalizes text and converts them to sequences of integers
18
+ 3) computes spectrograms from audio files.
19
+ """
20
+ def __init__(self, audiopaths_and_text, hparams):
21
+ self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
22
+ self.text_cleaners = hparams.text_cleaners
23
+ self.max_wav_value = hparams.max_wav_value
24
+ self.sampling_rate = hparams.sampling_rate
25
+ self.filter_length = hparams.filter_length
26
+ self.hop_length = hparams.hop_length
27
+ self.win_length = hparams.win_length
28
+ self.sampling_rate = hparams.sampling_rate
29
+
30
+ self.cleaned_text = getattr(hparams, "cleaned_text", False)
31
+
32
+ self.add_blank = hparams.add_blank
33
+ self.min_text_len = getattr(hparams, "min_text_len", 1)
34
+ self.max_text_len = getattr(hparams, "max_text_len", 190)
35
+
36
+ random.seed(1234)
37
+ random.shuffle(self.audiopaths_and_text)
38
+ self._filter()
39
+
40
+
41
+ def _filter(self):
42
+ """
43
+ Filter text & store spec lengths
44
+ """
45
+ # Store spectrogram lengths for Bucketing
46
+ # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
47
+ # spec_length = wav_length // hop_length
48
+
49
+ audiopaths_and_text_new = []
50
+ lengths = []
51
+ for audiopath, text in self.audiopaths_and_text:
52
+ if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
53
+ audiopaths_and_text_new.append([audiopath, text])
54
+ lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
55
+ self.audiopaths_and_text = audiopaths_and_text_new
56
+ self.lengths = lengths
57
+
58
+ def get_audio_text_pair(self, audiopath_and_text):
59
+ # separate filename and text
60
+ audiopath, text = audiopath_and_text[0], audiopath_and_text[1]
61
+ text = self.get_text(text)
62
+ spec, wav = self.get_audio(audiopath)
63
+ return (text, spec, wav)
64
+
65
+ def get_audio(self, filename):
66
+ audio, sampling_rate = load_wav_to_torch(filename)
67
+ if sampling_rate != self.sampling_rate:
68
+ raise ValueError("{} {} SR doesn't match target {} SR".format(
69
+ sampling_rate, self.sampling_rate))
70
+ audio_norm = audio / self.max_wav_value
71
+ audio_norm = audio_norm.unsqueeze(0)
72
+ spec_filename = filename.replace(".wav", ".spec.pt")
73
+ if os.path.exists(spec_filename):
74
+ spec = torch.load(spec_filename)
75
+ else:
76
+ spec = spectrogram_torch(audio_norm, self.filter_length,
77
+ self.sampling_rate, self.hop_length, self.win_length,
78
+ center=False)
79
+ spec = torch.squeeze(spec, 0)
80
+ torch.save(spec, spec_filename)
81
+ return spec, audio_norm
82
+
83
+ def get_text(self, text):
84
+ if self.cleaned_text:
85
+ text_norm = cleaned_text_to_sequence(text)
86
+ else:
87
+ text_norm = text_to_sequence(text, self.text_cleaners)
88
+ if self.add_blank:
89
+ text_norm = commons.intersperse(text_norm, 0)
90
+ text_norm = torch.LongTensor(text_norm)
91
+ return text_norm
92
+
93
+ def __getitem__(self, index):
94
+ return self.get_audio_text_pair(self.audiopaths_and_text[index])
95
+
96
+ def __len__(self):
97
+ return len(self.audiopaths_and_text)
98
+
99
+
100
+ class TextAudioCollate():
101
+ """ Zero-pads model inputs and targets
102
+ """
103
+ def __init__(self, return_ids=False):
104
+ self.return_ids = return_ids
105
+
106
+ def __call__(self, batch):
107
+ """Collate's training batch from normalized text and aduio
108
+ PARAMS
109
+ ------
110
+ batch: [text_normalized, spec_normalized, wav_normalized]
111
+ """
112
+ # Right zero-pad all one-hot text sequences to max input length
113
+ _, ids_sorted_decreasing = torch.sort(
114
+ torch.LongTensor([x[1].size(1) for x in batch]),
115
+ dim=0, descending=True)
116
+
117
+ max_text_len = max([len(x[0]) for x in batch])
118
+ max_spec_len = max([x[1].size(1) for x in batch])
119
+ max_wav_len = max([x[2].size(1) for x in batch])
120
+
121
+ text_lengths = torch.LongTensor(len(batch))
122
+ spec_lengths = torch.LongTensor(len(batch))
123
+ wav_lengths = torch.LongTensor(len(batch))
124
+
125
+ text_padded = torch.LongTensor(len(batch), max_text_len)
126
+ spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
127
+ wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
128
+ text_padded.zero_()
129
+ spec_padded.zero_()
130
+ wav_padded.zero_()
131
+ for i in range(len(ids_sorted_decreasing)):
132
+ row = batch[ids_sorted_decreasing[i]]
133
+
134
+ text = row[0]
135
+ text_padded[i, :text.size(0)] = text
136
+ text_lengths[i] = text.size(0)
137
+
138
+ spec = row[1]
139
+ spec_padded[i, :, :spec.size(1)] = spec
140
+ spec_lengths[i] = spec.size(1)
141
+
142
+ wav = row[2]
143
+ wav_padded[i, :, :wav.size(1)] = wav
144
+ wav_lengths[i] = wav.size(1)
145
+
146
+ if self.return_ids:
147
+ return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, ids_sorted_decreasing
148
+ return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths
149
+
150
+
151
+ """Multi speaker version"""
152
+ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
153
+ """
154
+ 1) loads audio, speaker_id, text pairs
155
+ 2) normalizes text and converts them to sequences of integers
156
+ 3) computes spectrograms from audio files.
157
+ """
158
+ def __init__(self, audiopaths_sid_text, hparams):
159
+ self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
160
+ self.text_cleaners = hparams.text_cleaners
161
+ self.max_wav_value = hparams.max_wav_value
162
+ self.sampling_rate = hparams.sampling_rate
163
+ self.filter_length = hparams.filter_length
164
+ self.hop_length = hparams.hop_length
165
+ self.win_length = hparams.win_length
166
+ self.sampling_rate = hparams.sampling_rate
167
+
168
+ self.cleaned_text = getattr(hparams, "cleaned_text", False)
169
+
170
+ self.add_blank = hparams.add_blank
171
+ self.min_text_len = getattr(hparams, "min_text_len", 1)
172
+ self.max_text_len = getattr(hparams, "max_text_len", 190)
173
+
174
+ random.seed(1234)
175
+ random.shuffle(self.audiopaths_sid_text)
176
+ self._filter()
177
+
178
+ def _filter(self):
179
+ """
180
+ Filter text & store spec lengths
181
+ """
182
+ # Store spectrogram lengths for Bucketing
183
+ # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
184
+ # spec_length = wav_length // hop_length
185
+
186
+ audiopaths_sid_text_new = []
187
+ lengths = []
188
+ for audiopath, sid, text in self.audiopaths_sid_text:
189
+ if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
190
+ audiopaths_sid_text_new.append([audiopath, sid, text])
191
+ lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
192
+ self.audiopaths_sid_text = audiopaths_sid_text_new
193
+ self.lengths = lengths
194
+
195
+ def get_audio_text_speaker_pair(self, audiopath_sid_text):
196
+ # separate filename, speaker_id and text
197
+ audiopath, sid, text = audiopath_sid_text[0], audiopath_sid_text[1], audiopath_sid_text[2]
198
+ text = self.get_text(text)
199
+ spec, wav = self.get_audio(audiopath)
200
+ sid = self.get_sid(sid)
201
+ return (text, spec, wav, sid)
202
+
203
+ def get_audio(self, filename):
204
+ audio, sampling_rate = load_wav_to_torch(filename)
205
+ if sampling_rate != self.sampling_rate:
206
+ raise ValueError("{} {} SR doesn't match target {} SR".format(
207
+ sampling_rate, self.sampling_rate))
208
+ audio_norm = audio / self.max_wav_value
209
+ audio_norm = audio_norm.unsqueeze(0)
210
+ spec_filename = filename.replace(".wav", ".spec.pt")
211
+ if os.path.exists(spec_filename):
212
+ spec = torch.load(spec_filename)
213
+ else:
214
+ spec = spectrogram_torch(audio_norm, self.filter_length,
215
+ self.sampling_rate, self.hop_length, self.win_length,
216
+ center=False)
217
+ spec = torch.squeeze(spec, 0)
218
+ torch.save(spec, spec_filename)
219
+ return spec, audio_norm
220
+
221
+ def get_text(self, text):
222
+ if self.cleaned_text:
223
+ text_norm = cleaned_text_to_sequence(text)
224
+ else:
225
+ text_norm = text_to_sequence(text, self.text_cleaners)
226
+ if self.add_blank:
227
+ text_norm = commons.intersperse(text_norm, 0)
228
+ text_norm = torch.LongTensor(text_norm)
229
+ return text_norm
230
+
231
+ def get_sid(self, sid):
232
+ sid = torch.LongTensor([int(sid)])
233
+ return sid
234
+
235
+ def __getitem__(self, index):
236
+ return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index])
237
+
238
+ def __len__(self):
239
+ return len(self.audiopaths_sid_text)
240
+
241
+
242
+ class TextAudioSpeakerCollate():
243
+ """ Zero-pads model inputs and targets
244
+ """
245
+ def __init__(self, return_ids=False):
246
+ self.return_ids = return_ids
247
+
248
+ def __call__(self, batch):
249
+ """Collate's training batch from normalized text, audio and speaker identities
250
+ PARAMS
251
+ ------
252
+ batch: [text_normalized, spec_normalized, wav_normalized, sid]
253
+ """
254
+ # Right zero-pad all one-hot text sequences to max input length
255
+ _, ids_sorted_decreasing = torch.sort(
256
+ torch.LongTensor([x[1].size(1) for x in batch]),
257
+ dim=0, descending=True)
258
+
259
+ max_text_len = max([len(x[0]) for x in batch])
260
+ max_spec_len = max([x[1].size(1) for x in batch])
261
+ max_wav_len = max([x[2].size(1) for x in batch])
262
+
263
+ text_lengths = torch.LongTensor(len(batch))
264
+ spec_lengths = torch.LongTensor(len(batch))
265
+ wav_lengths = torch.LongTensor(len(batch))
266
+ sid = torch.LongTensor(len(batch))
267
+
268
+ text_padded = torch.LongTensor(len(batch), max_text_len)
269
+ spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
270
+ wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
271
+ text_padded.zero_()
272
+ spec_padded.zero_()
273
+ wav_padded.zero_()
274
+ for i in range(len(ids_sorted_decreasing)):
275
+ row = batch[ids_sorted_decreasing[i]]
276
+
277
+ text = row[0]
278
+ text_padded[i, :text.size(0)] = text
279
+ text_lengths[i] = text.size(0)
280
+
281
+ spec = row[1]
282
+ spec_padded[i, :, :spec.size(1)] = spec
283
+ spec_lengths[i] = spec.size(1)
284
+
285
+ wav = row[2]
286
+ wav_padded[i, :, :wav.size(1)] = wav
287
+ wav_lengths[i] = wav.size(1)
288
+
289
+ sid[i] = row[3]
290
+
291
+ if self.return_ids:
292
+ return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, ids_sorted_decreasing
293
+ return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid
294
+
295
+
296
+ class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
297
+ """
298
+ Maintain similar input lengths in a batch.
299
+ Length groups are specified by boundaries.
300
+ Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
301
+
302
+ It removes samples which are not included in the boundaries.
303
+ Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
304
+ """
305
+ def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
306
+ super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
307
+ self.lengths = dataset.lengths
308
+ self.batch_size = batch_size
309
+ self.boundaries = boundaries
310
+
311
+ self.buckets, self.num_samples_per_bucket = self._create_buckets()
312
+ self.total_size = sum(self.num_samples_per_bucket)
313
+ self.num_samples = self.total_size // self.num_replicas
314
+
315
+ def _create_buckets(self):
316
+ buckets = [[] for _ in range(len(self.boundaries) - 1)]
317
+ for i in range(len(self.lengths)):
318
+ length = self.lengths[i]
319
+ idx_bucket = self._bisect(length)
320
+ if idx_bucket != -1:
321
+ buckets[idx_bucket].append(i)
322
+
323
+ for i in range(len(buckets) - 1, 0, -1):
324
+ if len(buckets[i]) == 0:
325
+ buckets.pop(i)
326
+ self.boundaries.pop(i+1)
327
+
328
+ num_samples_per_bucket = []
329
+ for i in range(len(buckets)):
330
+ len_bucket = len(buckets[i])
331
+ total_batch_size = self.num_replicas * self.batch_size
332
+ rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
333
+ num_samples_per_bucket.append(len_bucket + rem)
334
+ return buckets, num_samples_per_bucket
335
+
336
+ def __iter__(self):
337
+ # deterministically shuffle based on epoch
338
+ g = torch.Generator()
339
+ g.manual_seed(self.epoch)
340
+
341
+ indices = []
342
+ if self.shuffle:
343
+ for bucket in self.buckets:
344
+ indices.append(torch.randperm(len(bucket), generator=g).tolist())
345
+ else:
346
+ for bucket in self.buckets:
347
+ indices.append(list(range(len(bucket))))
348
+
349
+ batches = []
350
+ for i in range(len(self.buckets)):
351
+ bucket = self.buckets[i]
352
+ len_bucket = len(bucket)
353
+ ids_bucket = indices[i]
354
+ num_samples_bucket = self.num_samples_per_bucket[i]
355
+
356
+ # add extra samples to make it evenly divisible
357
+ rem = num_samples_bucket - len_bucket
358
+ ids_bucket = ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[:(rem % len_bucket)]
359
+
360
+ # subsample
361
+ ids_bucket = ids_bucket[self.rank::self.num_replicas]
362
+
363
+ # batching
364
+ for j in range(len(ids_bucket) // self.batch_size):
365
+ batch = [bucket[idx] for idx in ids_bucket[j*self.batch_size:(j+1)*self.batch_size]]
366
+ batches.append(batch)
367
+
368
+ if self.shuffle:
369
+ batch_ids = torch.randperm(len(batches), generator=g).tolist()
370
+ batches = [batches[i] for i in batch_ids]
371
+ self.batches = batches
372
+
373
+ assert len(self.batches) * self.batch_size == self.num_samples
374
+ return iter(self.batches)
375
+
376
+ def _bisect(self, x, lo=0, hi=None):
377
+ if hi is None:
378
+ hi = len(self.boundaries) - 1
379
+
380
+ if hi > lo:
381
+ mid = (hi + lo) // 2
382
+ if self.boundaries[mid] < x and x <= self.boundaries[mid+1]:
383
+ return mid
384
+ elif x <= self.boundaries[mid]:
385
+ return self._bisect(x, lo, mid)
386
+ else:
387
+ return self._bisect(x, mid + 1, hi)
388
+ else:
389
+ return -1
390
+
391
+ def __len__(self):
392
+ return self.num_samples // self.batch_size
inference_api.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import commons
3
+ import utils
4
+ from models import SynthesizerTrn
5
+ from text.symbols import symbols
6
+ from text import text_to_sequence
7
+ import io
8
+ from scipy.io.wavfile import write
9
+
10
+ from flask import Flask, request
11
+ import threading
12
+ app = Flask(__name__)
13
+ mutex = threading.Lock()
14
+
15
+ def get_text(text, hps):
16
+ text_norm = text_to_sequence(text, hps.data.text_cleaners)
17
+ if hps.data.add_blank:
18
+ text_norm = commons.intersperse(text_norm, 0)
19
+ text_norm = torch.LongTensor(text_norm)
20
+ return text_norm
21
+ hps = utils.get_hparams_from_file("./configs/ljs_mb_istft_vits.json")
22
+ net_g = SynthesizerTrn(
23
+ len(symbols),
24
+ hps.data.filter_length // 2 + 1,
25
+ hps.train.segment_size // hps.data.hop_length,
26
+ **hps.model)
27
+ _ = net_g.eval()
28
+
29
+ # _ = utils.load_checkpoint("../tempbest.pth", net_g, None)
30
+ import time
31
+
32
+
33
+ def tts(txt):
34
+ audio = None
35
+ if mutex.acquire(blocking=False):
36
+ try:
37
+ stn_tst = get_text(txt, hps)
38
+ with torch.no_grad():
39
+ x_tst = stn_tst.unsqueeze(0)
40
+ x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
41
+ t1 = time.time()
42
+ audio = net_g.infer(x_tst, x_tst_lengths, noise_scale=.667, noise_scale_w=0.8,
43
+ length_scale=1)[0][0, 0].data.float().numpy()
44
+ t2 = time.time()
45
+ print("推理时间:", (t2 - t1), "s")
46
+ finally:
47
+ mutex.release()
48
+ return audio
49
+
50
+ @app.route('/tts')
51
+ def text_api():
52
+ text = request.args.get('text','')
53
+ bytes_wav = bytes()
54
+ byte_io = io.BytesIO(bytes_wav)
55
+ audio = tts(text)
56
+ if audio is None:
57
+ return "服务器忙"
58
+ write(byte_io, 22050, audio)
59
+ wav_bytes = byte_io.read()
60
+
61
+ # audio_data = base64.b64encode(wav_bytes).decode('UTF-8')
62
+ return wav_bytes, 200, {'Content-Type': 'audio/wav'}
63
+
64
+
65
+ if __name__ == '__main__':
66
+ app.run("0.0.0.0", 8080)
losses.py ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+ from stft_loss import MultiResolutionSTFTLoss
4
+
5
+
6
+ import commons
7
+
8
+
9
+ def feature_loss(fmap_r, fmap_g):
10
+ loss = 0
11
+ for dr, dg in zip(fmap_r, fmap_g):
12
+ for rl, gl in zip(dr, dg):
13
+ rl = rl.float().detach()
14
+ gl = gl.float()
15
+ loss += torch.mean(torch.abs(rl - gl))
16
+
17
+ return loss * 2
18
+
19
+
20
+ def discriminator_loss(disc_real_outputs, disc_generated_outputs):
21
+ loss = 0
22
+ r_losses = []
23
+ g_losses = []
24
+ for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
25
+ dr = dr.float()
26
+ dg = dg.float()
27
+ r_loss = torch.mean((1-dr)**2)
28
+ g_loss = torch.mean(dg**2)
29
+ loss += (r_loss + g_loss)
30
+ r_losses.append(r_loss.item())
31
+ g_losses.append(g_loss.item())
32
+
33
+ return loss, r_losses, g_losses
34
+
35
+
36
+ def generator_loss(disc_outputs):
37
+ loss = 0
38
+ gen_losses = []
39
+ for dg in disc_outputs:
40
+ dg = dg.float()
41
+ l = torch.mean((1-dg)**2)
42
+ gen_losses.append(l)
43
+ loss += l
44
+
45
+ return loss, gen_losses
46
+
47
+
48
+ def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
49
+ """
50
+ z_p, logs_q: [b, h, t_t]
51
+ m_p, logs_p: [b, h, t_t]
52
+ """
53
+ z_p = z_p.float()
54
+ logs_q = logs_q.float()
55
+ m_p = m_p.float()
56
+ logs_p = logs_p.float()
57
+ z_mask = z_mask.float()
58
+
59
+ kl = logs_p - logs_q - 0.5
60
+ kl += 0.5 * ((z_p - m_p)**2) * torch.exp(-2. * logs_p)
61
+ kl = torch.sum(kl * z_mask)
62
+ l = kl / torch.sum(z_mask)
63
+ return l
64
+
65
+ def subband_stft_loss(h, y_mb, y_hat_mb):
66
+ sub_stft_loss = MultiResolutionSTFTLoss(h.train.fft_sizes, h.train.hop_sizes, h.train.win_lengths)
67
+ y_mb = y_mb.view(-1, y_mb.size(2))
68
+ y_hat_mb = y_hat_mb.view(-1, y_hat_mb.size(2))
69
+ sub_sc_loss, sub_mag_loss = sub_stft_loss(y_hat_mb[:, :y_mb.size(-1)], y_mb)
70
+ return sub_sc_loss+sub_mag_loss
71
+
mel_processing.py ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import os
3
+ import random
4
+ import torch
5
+ from torch import nn
6
+ import torch.nn.functional as F
7
+ import torch.utils.data
8
+ import numpy as np
9
+ import librosa
10
+ import librosa.util as librosa_util
11
+ from librosa.util import normalize, pad_center, tiny
12
+ from scipy.signal import get_window
13
+ from scipy.io.wavfile import read
14
+ from librosa.filters import mel as librosa_mel_fn
15
+
16
+ MAX_WAV_VALUE = 32768.0
17
+
18
+
19
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
20
+ """
21
+ PARAMS
22
+ ------
23
+ C: compression factor
24
+ """
25
+ return torch.log(torch.clamp(x, min=clip_val) * C)
26
+
27
+
28
+ def dynamic_range_decompression_torch(x, C=1):
29
+ """
30
+ PARAMS
31
+ ------
32
+ C: compression factor used to compress
33
+ """
34
+ return torch.exp(x) / C
35
+
36
+
37
+ def spectral_normalize_torch(magnitudes):
38
+ output = dynamic_range_compression_torch(magnitudes)
39
+ return output
40
+
41
+
42
+ def spectral_de_normalize_torch(magnitudes):
43
+ output = dynamic_range_decompression_torch(magnitudes)
44
+ return output
45
+
46
+
47
+ mel_basis = {}
48
+ hann_window = {}
49
+
50
+
51
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
52
+ if torch.min(y) < -1.:
53
+ print('min value is ', torch.min(y))
54
+ if torch.max(y) > 1.:
55
+ print('max value is ', torch.max(y))
56
+
57
+ global hann_window
58
+ dtype_device = str(y.dtype) + '_' + str(y.device)
59
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
60
+ if wnsize_dtype_device not in hann_window:
61
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
62
+
63
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
64
+ y = y.squeeze(1)
65
+
66
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
67
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
68
+
69
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
70
+ return spec
71
+
72
+
73
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
74
+ global mel_basis
75
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
76
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
77
+ if fmax_dtype_device not in mel_basis:
78
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
79
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
80
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
81
+ spec = spectral_normalize_torch(spec)
82
+ return spec
83
+
84
+
85
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
86
+ if torch.min(y) < -1.:
87
+ print('min value is ', torch.min(y))
88
+ if torch.max(y) > 1.:
89
+ print('max value is ', torch.max(y))
90
+
91
+ global mel_basis, hann_window
92
+ dtype_device = str(y.dtype) + '_' + str(y.device)
93
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
94
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
95
+ if fmax_dtype_device not in mel_basis:
96
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
97
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
98
+ if wnsize_dtype_device not in hann_window:
99
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
100
+
101
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
102
+ y = y.squeeze(1)
103
+
104
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
105
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
106
+
107
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
108
+
109
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
110
+ spec = spectral_normalize_torch(spec)
111
+
112
+ return spec
models.py ADDED
@@ -0,0 +1,730 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ import commons
8
+ import modules
9
+ import attentions
10
+ import monotonic_align
11
+
12
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
13
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
14
+ from commons import init_weights, get_padding
15
+ from pqmf import PQMF
16
+ from stft import TorchSTFT
17
+ import math
18
+
19
+
20
+ class StochasticDurationPredictor(nn.Module):
21
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
22
+ super().__init__()
23
+ filter_channels = in_channels # it needs to be removed from future version.
24
+ self.in_channels = in_channels
25
+ self.filter_channels = filter_channels
26
+ self.kernel_size = kernel_size
27
+ self.p_dropout = p_dropout
28
+ self.n_flows = n_flows
29
+ self.gin_channels = gin_channels
30
+
31
+ self.log_flow = modules.Log()
32
+ self.flows = nn.ModuleList()
33
+ self.flows.append(modules.ElementwiseAffine(2))
34
+ for i in range(n_flows):
35
+ self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
36
+ self.flows.append(modules.Flip())
37
+
38
+ self.post_pre = nn.Conv1d(1, filter_channels, 1)
39
+ self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
40
+ self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
41
+ self.post_flows = nn.ModuleList()
42
+ self.post_flows.append(modules.ElementwiseAffine(2))
43
+ for i in range(4):
44
+ self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
45
+ self.post_flows.append(modules.Flip())
46
+
47
+ self.pre = nn.Conv1d(in_channels, filter_channels, 1)
48
+ self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
49
+ self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
50
+ if gin_channels != 0:
51
+ self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
52
+
53
+ def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
54
+ x = torch.detach(x)
55
+ x = self.pre(x)
56
+ if g is not None:
57
+ g = torch.detach(g)
58
+ x = x + self.cond(g)
59
+ x = self.convs(x, x_mask)
60
+ x = self.proj(x) * x_mask
61
+
62
+ if not reverse:
63
+ flows = self.flows
64
+ assert w is not None
65
+
66
+ logdet_tot_q = 0
67
+ h_w = self.post_pre(w)
68
+ h_w = self.post_convs(h_w, x_mask)
69
+ h_w = self.post_proj(h_w) * x_mask
70
+ e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
71
+ z_q = e_q
72
+ for flow in self.post_flows:
73
+ z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
74
+ logdet_tot_q += logdet_q
75
+ z_u, z1 = torch.split(z_q, [1, 1], 1)
76
+ u = torch.sigmoid(z_u) * x_mask
77
+ z0 = (w - u) * x_mask
78
+ logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])
79
+ logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q
80
+
81
+ logdet_tot = 0
82
+ z0, logdet = self.log_flow(z0, x_mask)
83
+ logdet_tot += logdet
84
+ z = torch.cat([z0, z1], 1)
85
+ for flow in flows:
86
+ z, logdet = flow(z, x_mask, g=x, reverse=reverse)
87
+ logdet_tot = logdet_tot + logdet
88
+ nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot
89
+ return nll + logq # [b]
90
+ else:
91
+ flows = list(reversed(self.flows))
92
+ flows = flows[:-2] + [flows[-1]] # remove a useless vflow
93
+ z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
94
+ for flow in flows:
95
+ z = flow(z, x_mask, g=x, reverse=reverse)
96
+ z0, z1 = torch.split(z, [1, 1], 1)
97
+ logw = z0
98
+ return logw
99
+
100
+
101
+ class DurationPredictor(nn.Module):
102
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
103
+ super().__init__()
104
+
105
+ self.in_channels = in_channels
106
+ self.filter_channels = filter_channels
107
+ self.kernel_size = kernel_size
108
+ self.p_dropout = p_dropout
109
+ self.gin_channels = gin_channels
110
+
111
+ self.drop = nn.Dropout(p_dropout)
112
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)
113
+ self.norm_1 = modules.LayerNorm(filter_channels)
114
+ self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
115
+ self.norm_2 = modules.LayerNorm(filter_channels)
116
+ self.proj = nn.Conv1d(filter_channels, 1, 1)
117
+
118
+ if gin_channels != 0:
119
+ self.cond = nn.Conv1d(gin_channels, in_channels, 1)
120
+
121
+ def forward(self, x, x_mask, g=None):
122
+ x = torch.detach(x)
123
+ if g is not None:
124
+ g = torch.detach(g)
125
+ x = x + self.cond(g)
126
+ x = self.conv_1(x * x_mask)
127
+ x = torch.relu(x)
128
+ x = self.norm_1(x)
129
+ x = self.drop(x)
130
+ x = self.conv_2(x * x_mask)
131
+ x = torch.relu(x)
132
+ x = self.norm_2(x)
133
+ x = self.drop(x)
134
+ x = self.proj(x * x_mask)
135
+ return x * x_mask
136
+
137
+
138
+ class TextEncoder(nn.Module):
139
+ def __init__(self,
140
+ n_vocab,
141
+ out_channels,
142
+ hidden_channels,
143
+ filter_channels,
144
+ n_heads,
145
+ n_layers,
146
+ kernel_size,
147
+ p_dropout):
148
+ super().__init__()
149
+ self.n_vocab = n_vocab
150
+ self.out_channels = out_channels
151
+ self.hidden_channels = hidden_channels
152
+ self.filter_channels = filter_channels
153
+ self.n_heads = n_heads
154
+ self.n_layers = n_layers
155
+ self.kernel_size = kernel_size
156
+ self.p_dropout = p_dropout
157
+
158
+ self.emb = nn.Embedding(n_vocab, hidden_channels)
159
+ nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
160
+
161
+ self.encoder = attentions.Encoder(
162
+ hidden_channels,
163
+ filter_channels,
164
+ n_heads,
165
+ n_layers,
166
+ kernel_size,
167
+ p_dropout)
168
+ self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
169
+
170
+ def forward(self, x, x_lengths):
171
+ x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
172
+ x = torch.transpose(x, 1, -1) # [b, h, t]
173
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
174
+
175
+ x = self.encoder(x * x_mask, x_mask)
176
+ stats = self.proj(x) * x_mask
177
+
178
+ m, logs = torch.split(stats, self.out_channels, dim=1)
179
+ return x, m, logs, x_mask
180
+
181
+
182
+ class ResidualCouplingBlock(nn.Module):
183
+ def __init__(self,
184
+ channels,
185
+ hidden_channels,
186
+ kernel_size,
187
+ dilation_rate,
188
+ n_layers,
189
+ n_flows=4,
190
+ gin_channels=0):
191
+ super().__init__()
192
+ self.channels = channels
193
+ self.hidden_channels = hidden_channels
194
+ self.kernel_size = kernel_size
195
+ self.dilation_rate = dilation_rate
196
+ self.n_layers = n_layers
197
+ self.n_flows = n_flows
198
+ self.gin_channels = gin_channels
199
+
200
+ self.flows = nn.ModuleList()
201
+ for i in range(n_flows):
202
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
203
+ self.flows.append(modules.Flip())
204
+
205
+ def forward(self, x, x_mask, g=None, reverse=False):
206
+ if not reverse:
207
+ for flow in self.flows:
208
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
209
+ else:
210
+ for flow in reversed(self.flows):
211
+ x = flow(x, x_mask, g=g, reverse=reverse)
212
+ return x
213
+
214
+
215
+ class PosteriorEncoder(nn.Module):
216
+ def __init__(self,
217
+ in_channels,
218
+ out_channels,
219
+ hidden_channels,
220
+ kernel_size,
221
+ dilation_rate,
222
+ n_layers,
223
+ gin_channels=0):
224
+ super().__init__()
225
+ self.in_channels = in_channels
226
+ self.out_channels = out_channels
227
+ self.hidden_channels = hidden_channels
228
+ self.kernel_size = kernel_size
229
+ self.dilation_rate = dilation_rate
230
+ self.n_layers = n_layers
231
+ self.gin_channels = gin_channels
232
+
233
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
234
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
235
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
236
+
237
+ def forward(self, x, x_lengths, g=None):
238
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
239
+ x = self.pre(x) * x_mask
240
+ x = self.enc(x, x_mask, g=g)
241
+ stats = self.proj(x) * x_mask
242
+ m, logs = torch.split(stats, self.out_channels, dim=1)
243
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
244
+ return z, m, logs, x_mask
245
+
246
+ class iSTFT_Generator(torch.nn.Module):
247
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, gin_channels=0):
248
+ super(iSTFT_Generator, self).__init__()
249
+ # self.h = h
250
+ self.gen_istft_n_fft = gen_istft_n_fft
251
+ self.gen_istft_hop_size = gen_istft_hop_size
252
+
253
+ self.num_kernels = len(resblock_kernel_sizes)
254
+ self.num_upsamples = len(upsample_rates)
255
+ self.conv_pre = weight_norm(Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3))
256
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
257
+
258
+ self.ups = nn.ModuleList()
259
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
260
+ self.ups.append(weight_norm(
261
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
262
+ k, u, padding=(k-u)//2)))
263
+
264
+ self.resblocks = nn.ModuleList()
265
+ for i in range(len(self.ups)):
266
+ ch = upsample_initial_channel//(2**(i+1))
267
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
268
+ self.resblocks.append(resblock(ch, k, d))
269
+
270
+ self.post_n_fft = self.gen_istft_n_fft
271
+ self.conv_post = weight_norm(Conv1d(ch, self.post_n_fft + 2, 7, 1, padding=3))
272
+ self.ups.apply(init_weights)
273
+ self.conv_post.apply(init_weights)
274
+ self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
275
+ self.stft = TorchSTFT(filter_length=self.gen_istft_n_fft, hop_length=self.gen_istft_hop_size, win_length=self.gen_istft_n_fft)
276
+ def forward(self, x, g=None):
277
+
278
+ x = self.conv_pre(x)
279
+ for i in range(self.num_upsamples):
280
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
281
+ x = self.ups[i](x)
282
+ xs = None
283
+ for j in range(self.num_kernels):
284
+ if xs is None:
285
+ xs = self.resblocks[i*self.num_kernels+j](x)
286
+ else:
287
+ xs += self.resblocks[i*self.num_kernels+j](x)
288
+ x = xs / self.num_kernels
289
+ x = F.leaky_relu(x)
290
+ x = self.reflection_pad(x)
291
+ x = self.conv_post(x)
292
+ spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
293
+ phase = math.pi*torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
294
+ out = self.stft.inverse(spec, phase).to(x.device)
295
+ return out, None
296
+
297
+ def remove_weight_norm(self):
298
+ print('Removing weight norm...')
299
+ for l in self.ups:
300
+ remove_weight_norm(l)
301
+ for l in self.resblocks:
302
+ l.remove_weight_norm()
303
+ remove_weight_norm(self.conv_pre)
304
+ remove_weight_norm(self.conv_post)
305
+
306
+
307
+ class Multiband_iSTFT_Generator(torch.nn.Module):
308
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, subbands, gin_channels=0):
309
+ super(Multiband_iSTFT_Generator, self).__init__()
310
+ # self.h = h
311
+ self.subbands = subbands
312
+ self.num_kernels = len(resblock_kernel_sizes)
313
+ self.num_upsamples = len(upsample_rates)
314
+ self.conv_pre = weight_norm(Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3))
315
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
316
+
317
+ self.ups = nn.ModuleList()
318
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
319
+ self.ups.append(weight_norm(
320
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
321
+ k, u, padding=(k-u)//2)))
322
+
323
+ self.resblocks = nn.ModuleList()
324
+ for i in range(len(self.ups)):
325
+ ch = upsample_initial_channel//(2**(i+1))
326
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
327
+ self.resblocks.append(resblock(ch, k, d))
328
+
329
+ self.post_n_fft = gen_istft_n_fft
330
+ self.ups.apply(init_weights)
331
+ self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
332
+ self.reshape_pixelshuffle = []
333
+
334
+ self.subband_conv_post = weight_norm(Conv1d(ch, self.subbands*(self.post_n_fft + 2), 7, 1, padding=3))
335
+
336
+ self.subband_conv_post.apply(init_weights)
337
+
338
+ self.gen_istft_n_fft = gen_istft_n_fft
339
+ self.gen_istft_hop_size = gen_istft_hop_size
340
+
341
+
342
+ def forward(self, x, g=None):
343
+ stft = TorchSTFT(filter_length=self.gen_istft_n_fft, hop_length=self.gen_istft_hop_size, win_length=self.gen_istft_n_fft).to(x.device)
344
+ pqmf = PQMF(x.device)
345
+
346
+ x = self.conv_pre(x)#[B, ch, length]
347
+
348
+ for i in range(self.num_upsamples):
349
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
350
+ x = self.ups[i](x)
351
+
352
+
353
+ xs = None
354
+ for j in range(self.num_kernels):
355
+ if xs is None:
356
+ xs = self.resblocks[i*self.num_kernels+j](x)
357
+ else:
358
+ xs += self.resblocks[i*self.num_kernels+j](x)
359
+ x = xs / self.num_kernels
360
+
361
+ x = F.leaky_relu(x)
362
+ x = self.reflection_pad(x)
363
+ x = self.subband_conv_post(x)
364
+ x = torch.reshape(x, (x.shape[0], self.subbands, x.shape[1]//self.subbands, x.shape[-1]))
365
+
366
+ spec = torch.exp(x[:,:,:self.post_n_fft // 2 + 1, :])
367
+ phase = math.pi*torch.sin(x[:,:, self.post_n_fft // 2 + 1:, :])
368
+
369
+ y_mb_hat = stft.inverse(torch.reshape(spec, (spec.shape[0]*self.subbands, self.gen_istft_n_fft // 2 + 1, spec.shape[-1])), torch.reshape(phase, (phase.shape[0]*self.subbands, self.gen_istft_n_fft // 2 + 1, phase.shape[-1])))
370
+ y_mb_hat = torch.reshape(y_mb_hat, (x.shape[0], self.subbands, 1, y_mb_hat.shape[-1]))
371
+ y_mb_hat = y_mb_hat.squeeze(-2)
372
+
373
+ y_g_hat = pqmf.synthesis(y_mb_hat)
374
+
375
+ return y_g_hat, y_mb_hat
376
+
377
+ def remove_weight_norm(self):
378
+ print('Removing weight norm...')
379
+ for l in self.ups:
380
+ remove_weight_norm(l)
381
+ for l in self.resblocks:
382
+ l.remove_weight_norm()
383
+
384
+
385
+ class Multistream_iSTFT_Generator(torch.nn.Module):
386
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, subbands, gin_channels=0):
387
+ super(Multistream_iSTFT_Generator, self).__init__()
388
+ # self.h = h
389
+ self.subbands = subbands
390
+ self.num_kernels = len(resblock_kernel_sizes)
391
+ self.num_upsamples = len(upsample_rates)
392
+ self.conv_pre = weight_norm(Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3))
393
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
394
+
395
+ self.ups = nn.ModuleList()
396
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
397
+ self.ups.append(weight_norm(
398
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
399
+ k, u, padding=(k-u)//2)))
400
+
401
+ self.resblocks = nn.ModuleList()
402
+ for i in range(len(self.ups)):
403
+ ch = upsample_initial_channel//(2**(i+1))
404
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
405
+ self.resblocks.append(resblock(ch, k, d))
406
+
407
+ self.post_n_fft = gen_istft_n_fft
408
+ self.ups.apply(init_weights)
409
+ self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
410
+ self.reshape_pixelshuffle = []
411
+
412
+ self.subband_conv_post = weight_norm(Conv1d(ch, self.subbands*(self.post_n_fft + 2), 7, 1, padding=3))
413
+
414
+ self.subband_conv_post.apply(init_weights)
415
+
416
+ self.gen_istft_n_fft = gen_istft_n_fft
417
+ self.gen_istft_hop_size = gen_istft_hop_size
418
+
419
+ updown_filter = torch.zeros((self.subbands, self.subbands, self.subbands)).float()
420
+ for k in range(self.subbands):
421
+ updown_filter[k, k, 0] = 1.0
422
+ self.register_buffer("updown_filter", updown_filter)
423
+ self.multistream_conv_post = weight_norm(Conv1d(4, 1, kernel_size=63, bias=False, padding=get_padding(63, 1)))
424
+ self.multistream_conv_post.apply(init_weights)
425
+
426
+
427
+
428
+ def forward(self, x, g=None):
429
+ stft = TorchSTFT(filter_length=self.gen_istft_n_fft, hop_length=self.gen_istft_hop_size, win_length=self.gen_istft_n_fft).to(x.device)
430
+ # pqmf = PQMF(x.device)
431
+
432
+ x = self.conv_pre(x)#[B, ch, length]
433
+
434
+ for i in range(self.num_upsamples):
435
+
436
+
437
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
438
+ x = self.ups[i](x)
439
+
440
+
441
+ xs = None
442
+ for j in range(self.num_kernels):
443
+ if xs is None:
444
+ xs = self.resblocks[i*self.num_kernels+j](x)
445
+ else:
446
+ xs += self.resblocks[i*self.num_kernels+j](x)
447
+ x = xs / self.num_kernels
448
+
449
+ x = F.leaky_relu(x)
450
+ x = self.reflection_pad(x)
451
+ x = self.subband_conv_post(x)
452
+ x = torch.reshape(x, (x.shape[0], self.subbands, x.shape[1]//self.subbands, x.shape[-1]))
453
+
454
+ spec = torch.exp(x[:,:,:self.post_n_fft // 2 + 1, :])
455
+ phase = math.pi*torch.sin(x[:,:, self.post_n_fft // 2 + 1:, :])
456
+
457
+ y_mb_hat = stft.inverse(torch.reshape(spec, (spec.shape[0]*self.subbands, self.gen_istft_n_fft // 2 + 1, spec.shape[-1])), torch.reshape(phase, (phase.shape[0]*self.subbands, self.gen_istft_n_fft // 2 + 1, phase.shape[-1])))
458
+ y_mb_hat = torch.reshape(y_mb_hat, (x.shape[0], self.subbands, 1, y_mb_hat.shape[-1]))
459
+ y_mb_hat = y_mb_hat.squeeze(-2)
460
+
461
+ y_mb_hat = F.conv_transpose1d(y_mb_hat, self.updown_filter.to(x.device) * self.subbands, stride=self.subbands)
462
+
463
+ y_g_hat = self.multistream_conv_post(y_mb_hat)
464
+
465
+ return y_g_hat, y_mb_hat
466
+
467
+ def remove_weight_norm(self):
468
+ print('Removing weight norm...')
469
+ for l in self.ups:
470
+ remove_weight_norm(l)
471
+ for l in self.resblocks:
472
+ l.remove_weight_norm()
473
+
474
+
475
+ class DiscriminatorP(torch.nn.Module):
476
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
477
+ super(DiscriminatorP, self).__init__()
478
+ self.period = period
479
+ self.use_spectral_norm = use_spectral_norm
480
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
481
+ self.convs = nn.ModuleList([
482
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
483
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
484
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
485
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
486
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
487
+ ])
488
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
489
+
490
+ def forward(self, x):
491
+ fmap = []
492
+
493
+ # 1d to 2d
494
+ b, c, t = x.shape
495
+ if t % self.period != 0: # pad first
496
+ n_pad = self.period - (t % self.period)
497
+ x = F.pad(x, (0, n_pad), "reflect")
498
+ t = t + n_pad
499
+ x = x.view(b, c, t // self.period, self.period)
500
+
501
+ for l in self.convs:
502
+ x = l(x)
503
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
504
+ fmap.append(x)
505
+ x = self.conv_post(x)
506
+ fmap.append(x)
507
+ x = torch.flatten(x, 1, -1)
508
+
509
+ return x, fmap
510
+
511
+
512
+ class DiscriminatorS(torch.nn.Module):
513
+ def __init__(self, use_spectral_norm=False):
514
+ super(DiscriminatorS, self).__init__()
515
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
516
+ self.convs = nn.ModuleList([
517
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
518
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
519
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
520
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
521
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
522
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
523
+ ])
524
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
525
+
526
+ def forward(self, x):
527
+ fmap = []
528
+
529
+ for l in self.convs:
530
+ x = l(x)
531
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
532
+ fmap.append(x)
533
+ x = self.conv_post(x)
534
+ fmap.append(x)
535
+ x = torch.flatten(x, 1, -1)
536
+
537
+ return x, fmap
538
+
539
+
540
+ class MultiPeriodDiscriminator(torch.nn.Module):
541
+ def __init__(self, use_spectral_norm=False):
542
+ super(MultiPeriodDiscriminator, self).__init__()
543
+ periods = [2,3,5,7,11]
544
+
545
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
546
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
547
+ self.discriminators = nn.ModuleList(discs)
548
+
549
+ def forward(self, y, y_hat):
550
+ y_d_rs = []
551
+ y_d_gs = []
552
+ fmap_rs = []
553
+ fmap_gs = []
554
+ for i, d in enumerate(self.discriminators):
555
+ y_d_r, fmap_r = d(y)
556
+ y_d_g, fmap_g = d(y_hat)
557
+ y_d_rs.append(y_d_r)
558
+ y_d_gs.append(y_d_g)
559
+ fmap_rs.append(fmap_r)
560
+ fmap_gs.append(fmap_g)
561
+
562
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
563
+
564
+
565
+
566
+ class SynthesizerTrn(nn.Module):
567
+ """
568
+ Synthesizer for Training
569
+ """
570
+
571
+ def __init__(self,
572
+ n_vocab,
573
+ spec_channels,
574
+ segment_size,
575
+ inter_channels,
576
+ hidden_channels,
577
+ filter_channels,
578
+ n_heads,
579
+ n_layers,
580
+ kernel_size,
581
+ p_dropout,
582
+ resblock,
583
+ resblock_kernel_sizes,
584
+ resblock_dilation_sizes,
585
+ upsample_rates,
586
+ upsample_initial_channel,
587
+ upsample_kernel_sizes,
588
+ gen_istft_n_fft,
589
+ gen_istft_hop_size,
590
+ n_speakers=0,
591
+ gin_channels=0,
592
+ use_sdp=False,
593
+ ms_istft_vits=False,
594
+ mb_istft_vits = False,
595
+ subbands = False,
596
+ istft_vits=False,
597
+ **kwargs):
598
+
599
+ super().__init__()
600
+ self.n_vocab = n_vocab
601
+ self.spec_channels = spec_channels
602
+ self.inter_channels = inter_channels
603
+ self.hidden_channels = hidden_channels
604
+ self.filter_channels = filter_channels
605
+ self.n_heads = n_heads
606
+ self.n_layers = n_layers
607
+ self.kernel_size = kernel_size
608
+ self.p_dropout = p_dropout
609
+ self.resblock = resblock
610
+ self.resblock_kernel_sizes = resblock_kernel_sizes
611
+ self.resblock_dilation_sizes = resblock_dilation_sizes
612
+ self.upsample_rates = upsample_rates
613
+ self.upsample_initial_channel = upsample_initial_channel
614
+ self.upsample_kernel_sizes = upsample_kernel_sizes
615
+ self.segment_size = segment_size
616
+ self.n_speakers = n_speakers
617
+ self.gin_channels = gin_channels
618
+ self.ms_istft_vits = ms_istft_vits
619
+ self.mb_istft_vits = mb_istft_vits
620
+ self.istft_vits = istft_vits
621
+
622
+ self.use_sdp = use_sdp
623
+
624
+ self.enc_p = TextEncoder(n_vocab,
625
+ inter_channels,
626
+ hidden_channels,
627
+ filter_channels,
628
+ n_heads,
629
+ n_layers,
630
+ kernel_size,
631
+ p_dropout)
632
+ if mb_istft_vits == True:
633
+ print('Mutli-band iSTFT VITS')
634
+ self.dec = Multiband_iSTFT_Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, subbands, gin_channels=gin_channels)
635
+ elif ms_istft_vits == True:
636
+ print('Mutli-stream iSTFT VITS')
637
+ self.dec = Multistream_iSTFT_Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, subbands, gin_channels=gin_channels)
638
+ elif istft_vits == True:
639
+ print('iSTFT-VITS')
640
+ self.dec = iSTFT_Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size, gin_channels=gin_channels)
641
+ else:
642
+ print('Decoder Error in json file')
643
+
644
+ self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
645
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
646
+
647
+ if use_sdp:
648
+ self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
649
+ else:
650
+ self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
651
+
652
+ if n_speakers > 1:
653
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
654
+
655
+ def forward(self, x, x_lengths, y, y_lengths, sid=None):
656
+
657
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
658
+ if self.n_speakers > 0:
659
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
660
+ else:
661
+ g = None
662
+
663
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
664
+ z_p = self.flow(z, y_mask, g=g)
665
+
666
+ with torch.no_grad():
667
+ # negative cross-entropy
668
+ s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
669
+ neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
670
+ neg_cent2 = torch.matmul(-0.5 * (z_p ** 2).transpose(1, 2), s_p_sq_r) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
671
+ neg_cent3 = torch.matmul(z_p.transpose(1, 2), (m_p * s_p_sq_r)) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
672
+ neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
673
+ neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
674
+
675
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
676
+ attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
677
+
678
+ w = attn.sum(2)
679
+ if self.use_sdp:
680
+ l_length = self.dp(x, x_mask, w, g=g)
681
+ l_length = l_length / torch.sum(x_mask)
682
+ else:
683
+ logw_ = torch.log(w + 1e-6) * x_mask
684
+ logw = self.dp(x, x_mask, g=g)
685
+ l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging
686
+
687
+ # expand prior
688
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
689
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
690
+
691
+ z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
692
+ o, o_mb = self.dec(z_slice, g=g)
693
+ return o, o_mb, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
694
+
695
+ def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
696
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
697
+ if self.n_speakers > 0:
698
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
699
+ else:
700
+ g = None
701
+
702
+ if self.use_sdp:
703
+ logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
704
+ else:
705
+ logw = self.dp(x, x_mask, g=g)
706
+ w = torch.exp(logw) * x_mask * length_scale
707
+ w_ceil = torch.ceil(w)
708
+ y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
709
+ y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
710
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
711
+ attn = commons.generate_path(w_ceil, attn_mask)
712
+
713
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
714
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
715
+
716
+ z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
717
+ z = self.flow(z_p, y_mask, g=g, reverse=True)
718
+ o, o_mb = self.dec((z * y_mask)[:,:,:max_len], g=g)
719
+ return o, o_mb, attn, y_mask, (z, z_p, m_p, logs_p)
720
+
721
+ def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
722
+ assert self.n_speakers > 0, "n_speakers have to be larger than 0."
723
+ g_src = self.emb_g(sid_src).unsqueeze(-1)
724
+ g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
725
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
726
+ z_p = self.flow(z, y_mask, g=g_src)
727
+ z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
728
+ o_hat, o_hat_mb = self.dec(z_hat * y_mask, g=g_tgt)
729
+ return o_hat, o_hat_mb, y_mask, (z, z_p, z_hat)
730
+
modules.py ADDED
@@ -0,0 +1,390 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import numpy as np
4
+ import scipy
5
+ import torch
6
+ from torch import nn
7
+ from torch.nn import functional as F
8
+
9
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
10
+ from torch.nn.utils import weight_norm, remove_weight_norm
11
+
12
+ import commons
13
+ from commons import init_weights, get_padding
14
+ from transforms import piecewise_rational_quadratic_transform
15
+
16
+
17
+ LRELU_SLOPE = 0.1
18
+
19
+
20
+ class LayerNorm(nn.Module):
21
+ def __init__(self, channels, eps=1e-5):
22
+ super().__init__()
23
+ self.channels = channels
24
+ self.eps = eps
25
+
26
+ self.gamma = nn.Parameter(torch.ones(channels))
27
+ self.beta = nn.Parameter(torch.zeros(channels))
28
+
29
+ def forward(self, x):
30
+ x = x.transpose(1, -1)
31
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
32
+ return x.transpose(1, -1)
33
+
34
+
35
+ class ConvReluNorm(nn.Module):
36
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
37
+ super().__init__()
38
+ self.in_channels = in_channels
39
+ self.hidden_channels = hidden_channels
40
+ self.out_channels = out_channels
41
+ self.kernel_size = kernel_size
42
+ self.n_layers = n_layers
43
+ self.p_dropout = p_dropout
44
+ assert n_layers > 1, "Number of layers should be larger than 0."
45
+
46
+ self.conv_layers = nn.ModuleList()
47
+ self.norm_layers = nn.ModuleList()
48
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
49
+ self.norm_layers.append(LayerNorm(hidden_channels))
50
+ self.relu_drop = nn.Sequential(
51
+ nn.ReLU(),
52
+ nn.Dropout(p_dropout))
53
+ for _ in range(n_layers-1):
54
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
55
+ self.norm_layers.append(LayerNorm(hidden_channels))
56
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
57
+ self.proj.weight.data.zero_()
58
+ self.proj.bias.data.zero_()
59
+
60
+ def forward(self, x, x_mask):
61
+ x_org = x
62
+ for i in range(self.n_layers):
63
+ x = self.conv_layers[i](x * x_mask)
64
+ x = self.norm_layers[i](x)
65
+ x = self.relu_drop(x)
66
+ x = x_org + self.proj(x)
67
+ return x * x_mask
68
+
69
+
70
+ class DDSConv(nn.Module):
71
+ """
72
+ Dialted and Depth-Separable Convolution
73
+ """
74
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
75
+ super().__init__()
76
+ self.channels = channels
77
+ self.kernel_size = kernel_size
78
+ self.n_layers = n_layers
79
+ self.p_dropout = p_dropout
80
+
81
+ self.drop = nn.Dropout(p_dropout)
82
+ self.convs_sep = nn.ModuleList()
83
+ self.convs_1x1 = nn.ModuleList()
84
+ self.norms_1 = nn.ModuleList()
85
+ self.norms_2 = nn.ModuleList()
86
+ for i in range(n_layers):
87
+ dilation = kernel_size ** i
88
+ padding = (kernel_size * dilation - dilation) // 2
89
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
90
+ groups=channels, dilation=dilation, padding=padding
91
+ ))
92
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
93
+ self.norms_1.append(LayerNorm(channels))
94
+ self.norms_2.append(LayerNorm(channels))
95
+
96
+ def forward(self, x, x_mask, g=None):
97
+ if g is not None:
98
+ x = x + g
99
+ for i in range(self.n_layers):
100
+ y = self.convs_sep[i](x * x_mask)
101
+ y = self.norms_1[i](y)
102
+ y = F.gelu(y)
103
+ y = self.convs_1x1[i](y)
104
+ y = self.norms_2[i](y)
105
+ y = F.gelu(y)
106
+ y = self.drop(y)
107
+ x = x + y
108
+ return x * x_mask
109
+
110
+
111
+ class WN(torch.nn.Module):
112
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
113
+ super(WN, self).__init__()
114
+ assert(kernel_size % 2 == 1)
115
+ self.hidden_channels =hidden_channels
116
+ self.kernel_size = kernel_size,
117
+ self.dilation_rate = dilation_rate
118
+ self.n_layers = n_layers
119
+ self.gin_channels = gin_channels
120
+ self.p_dropout = p_dropout
121
+
122
+ self.in_layers = torch.nn.ModuleList()
123
+ self.res_skip_layers = torch.nn.ModuleList()
124
+ self.drop = nn.Dropout(p_dropout)
125
+
126
+ if gin_channels != 0:
127
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
128
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
129
+
130
+ for i in range(n_layers):
131
+ dilation = dilation_rate ** i
132
+ padding = int((kernel_size * dilation - dilation) / 2)
133
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
134
+ dilation=dilation, padding=padding)
135
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
136
+ self.in_layers.append(in_layer)
137
+
138
+ # last one is not necessary
139
+ if i < n_layers - 1:
140
+ res_skip_channels = 2 * hidden_channels
141
+ else:
142
+ res_skip_channels = hidden_channels
143
+
144
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
145
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
146
+ self.res_skip_layers.append(res_skip_layer)
147
+
148
+ def forward(self, x, x_mask, g=None, **kwargs):
149
+ output = torch.zeros_like(x)
150
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
151
+
152
+ if g is not None:
153
+ g = self.cond_layer(g)
154
+
155
+ for i in range(self.n_layers):
156
+ x_in = self.in_layers[i](x)
157
+ if g is not None:
158
+ cond_offset = i * 2 * self.hidden_channels
159
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
160
+ else:
161
+ g_l = torch.zeros_like(x_in)
162
+
163
+ acts = commons.fused_add_tanh_sigmoid_multiply(
164
+ x_in,
165
+ g_l,
166
+ n_channels_tensor)
167
+ acts = self.drop(acts)
168
+
169
+ res_skip_acts = self.res_skip_layers[i](acts)
170
+ if i < self.n_layers - 1:
171
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
172
+ x = (x + res_acts) * x_mask
173
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
174
+ else:
175
+ output = output + res_skip_acts
176
+ return output * x_mask
177
+
178
+ def remove_weight_norm(self):
179
+ if self.gin_channels != 0:
180
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
181
+ for l in self.in_layers:
182
+ torch.nn.utils.remove_weight_norm(l)
183
+ for l in self.res_skip_layers:
184
+ torch.nn.utils.remove_weight_norm(l)
185
+
186
+
187
+ class ResBlock1(torch.nn.Module):
188
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
189
+ super(ResBlock1, self).__init__()
190
+ self.convs1 = nn.ModuleList([
191
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
192
+ padding=get_padding(kernel_size, dilation[0]))),
193
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
194
+ padding=get_padding(kernel_size, dilation[1]))),
195
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
196
+ padding=get_padding(kernel_size, dilation[2])))
197
+ ])
198
+ self.convs1.apply(init_weights)
199
+
200
+ self.convs2 = nn.ModuleList([
201
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
202
+ padding=get_padding(kernel_size, 1))),
203
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
204
+ padding=get_padding(kernel_size, 1))),
205
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
206
+ padding=get_padding(kernel_size, 1)))
207
+ ])
208
+ self.convs2.apply(init_weights)
209
+
210
+ def forward(self, x, x_mask=None):
211
+ for c1, c2 in zip(self.convs1, self.convs2):
212
+ xt = F.leaky_relu(x, LRELU_SLOPE)
213
+ if x_mask is not None:
214
+ xt = xt * x_mask
215
+ xt = c1(xt)
216
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
217
+ if x_mask is not None:
218
+ xt = xt * x_mask
219
+ xt = c2(xt)
220
+ x = xt + x
221
+ if x_mask is not None:
222
+ x = x * x_mask
223
+ return x
224
+
225
+ def remove_weight_norm(self):
226
+ for l in self.convs1:
227
+ remove_weight_norm(l)
228
+ for l in self.convs2:
229
+ remove_weight_norm(l)
230
+
231
+
232
+ class ResBlock2(torch.nn.Module):
233
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
234
+ super(ResBlock2, self).__init__()
235
+ self.convs = nn.ModuleList([
236
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
237
+ padding=get_padding(kernel_size, dilation[0]))),
238
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
239
+ padding=get_padding(kernel_size, dilation[1])))
240
+ ])
241
+ self.convs.apply(init_weights)
242
+
243
+ def forward(self, x, x_mask=None):
244
+ for c in self.convs:
245
+ xt = F.leaky_relu(x, LRELU_SLOPE)
246
+ if x_mask is not None:
247
+ xt = xt * x_mask
248
+ xt = c(xt)
249
+ x = xt + x
250
+ if x_mask is not None:
251
+ x = x * x_mask
252
+ return x
253
+
254
+ def remove_weight_norm(self):
255
+ for l in self.convs:
256
+ remove_weight_norm(l)
257
+
258
+
259
+ class Log(nn.Module):
260
+ def forward(self, x, x_mask, reverse=False, **kwargs):
261
+ if not reverse:
262
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
263
+ logdet = torch.sum(-y, [1, 2])
264
+ return y, logdet
265
+ else:
266
+ x = torch.exp(x) * x_mask
267
+ return x
268
+
269
+
270
+ class Flip(nn.Module):
271
+ def forward(self, x, *args, reverse=False, **kwargs):
272
+ x = torch.flip(x, [1])
273
+ if not reverse:
274
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
275
+ return x, logdet
276
+ else:
277
+ return x
278
+
279
+
280
+ class ElementwiseAffine(nn.Module):
281
+ def __init__(self, channels):
282
+ super().__init__()
283
+ self.channels = channels
284
+ self.m = nn.Parameter(torch.zeros(channels,1))
285
+ self.logs = nn.Parameter(torch.zeros(channels,1))
286
+
287
+ def forward(self, x, x_mask, reverse=False, **kwargs):
288
+ if not reverse:
289
+ y = self.m + torch.exp(self.logs) * x
290
+ y = y * x_mask
291
+ logdet = torch.sum(self.logs * x_mask, [1,2])
292
+ return y, logdet
293
+ else:
294
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
295
+ return x
296
+
297
+
298
+ class ResidualCouplingLayer(nn.Module):
299
+ def __init__(self,
300
+ channels,
301
+ hidden_channels,
302
+ kernel_size,
303
+ dilation_rate,
304
+ n_layers,
305
+ p_dropout=0,
306
+ gin_channels=0,
307
+ mean_only=False):
308
+ assert channels % 2 == 0, "channels should be divisible by 2"
309
+ super().__init__()
310
+ self.channels = channels
311
+ self.hidden_channels = hidden_channels
312
+ self.kernel_size = kernel_size
313
+ self.dilation_rate = dilation_rate
314
+ self.n_layers = n_layers
315
+ self.half_channels = channels // 2
316
+ self.mean_only = mean_only
317
+
318
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
319
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
320
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
321
+ self.post.weight.data.zero_()
322
+ self.post.bias.data.zero_()
323
+
324
+ def forward(self, x, x_mask, g=None, reverse=False):
325
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
326
+ h = self.pre(x0) * x_mask
327
+ h = self.enc(h, x_mask, g=g)
328
+ stats = self.post(h) * x_mask
329
+ if not self.mean_only:
330
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
331
+ else:
332
+ m = stats
333
+ logs = torch.zeros_like(m)
334
+
335
+ if not reverse:
336
+ x1 = m + x1 * torch.exp(logs) * x_mask
337
+ x = torch.cat([x0, x1], 1)
338
+ logdet = torch.sum(logs, [1,2])
339
+ return x, logdet
340
+ else:
341
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
342
+ x = torch.cat([x0, x1], 1)
343
+ return x
344
+
345
+
346
+ class ConvFlow(nn.Module):
347
+ def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
348
+ super().__init__()
349
+ self.in_channels = in_channels
350
+ self.filter_channels = filter_channels
351
+ self.kernel_size = kernel_size
352
+ self.n_layers = n_layers
353
+ self.num_bins = num_bins
354
+ self.tail_bound = tail_bound
355
+ self.half_channels = in_channels // 2
356
+
357
+ self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
358
+ self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
359
+ self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
360
+ self.proj.weight.data.zero_()
361
+ self.proj.bias.data.zero_()
362
+
363
+ def forward(self, x, x_mask, g=None, reverse=False):
364
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
365
+ h = self.pre(x0)
366
+ h = self.convs(h, x_mask, g=g)
367
+ h = self.proj(h) * x_mask
368
+
369
+ b, c, t = x0.shape
370
+ h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
371
+
372
+ unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
373
+ unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
374
+ unnormalized_derivatives = h[..., 2 * self.num_bins:]
375
+
376
+ x1, logabsdet = piecewise_rational_quadratic_transform(x1,
377
+ unnormalized_widths,
378
+ unnormalized_heights,
379
+ unnormalized_derivatives,
380
+ inverse=reverse,
381
+ tails='linear',
382
+ tail_bound=self.tail_bound
383
+ )
384
+
385
+ x = torch.cat([x0, x1], 1) * x_mask
386
+ logdet = torch.sum(logabsdet * x_mask, [1,2])
387
+ if not reverse:
388
+ return x, logdet
389
+ else:
390
+ return x
monotonic_align/__init__.py ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import torch
3
+ from .monotonic_align.core import maximum_path_c
4
+
5
+
6
+ def maximum_path(neg_cent, mask):
7
+ """ Cython optimized version.
8
+ neg_cent: [b, t_t, t_s]
9
+ mask: [b, t_t, t_s]
10
+ """
11
+ device = neg_cent.device
12
+ dtype = neg_cent.dtype
13
+ neg_cent = neg_cent.data.cpu().numpy().astype(np.float32)
14
+ path = np.zeros(neg_cent.shape, dtype=np.int32)
15
+
16
+ t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(np.int32)
17
+ t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(np.int32)
18
+ maximum_path_c(path, neg_cent, t_t_max, t_s_max)
19
+ return torch.from_numpy(path).to(device=device, dtype=dtype)
monotonic_align/core.pyx ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ cimport cython
2
+ from cython.parallel import prange
3
+
4
+
5
+ @cython.boundscheck(False)
6
+ @cython.wraparound(False)
7
+ cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_y, int t_x, float max_neg_val=-1e9) nogil:
8
+ cdef int x
9
+ cdef int y
10
+ cdef float v_prev
11
+ cdef float v_cur
12
+ cdef float tmp
13
+ cdef int index = t_x - 1
14
+
15
+ for y in range(t_y):
16
+ for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
17
+ if x == y:
18
+ v_cur = max_neg_val
19
+ else:
20
+ v_cur = value[y-1, x]
21
+ if x == 0:
22
+ if y == 0:
23
+ v_prev = 0.
24
+ else:
25
+ v_prev = max_neg_val
26
+ else:
27
+ v_prev = value[y-1, x-1]
28
+ value[y, x] += max(v_prev, v_cur)
29
+
30
+ for y in range(t_y - 1, -1, -1):
31
+ path[y, index] = 1
32
+ if index != 0 and (index == y or value[y-1, index] < value[y-1, index-1]):
33
+ index = index - 1
34
+
35
+
36
+ @cython.boundscheck(False)
37
+ @cython.wraparound(False)
38
+ cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_ys, int[::1] t_xs) nogil:
39
+ cdef int b = paths.shape[0]
40
+ cdef int i
41
+ for i in prange(b, nogil=True):
42
+ maximum_path_each(paths[i], values[i], t_ys[i], t_xs[i])
monotonic_align/setup.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ from distutils.core import setup
2
+ from Cython.Build import cythonize
3
+ import numpy
4
+
5
+ setup(
6
+ name = 'monotonic_align',
7
+ ext_modules = cythonize("core.pyx"),
8
+ include_dirs=[numpy.get_include()]
9
+ )
pqmf.py ADDED
@@ -0,0 +1,116 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+
3
+ # Copyright 2020 Tomoki Hayashi
4
+ # MIT License (https://opensource.org/licenses/MIT)
5
+
6
+ """Pseudo QMF modules."""
7
+
8
+ import numpy as np
9
+ import torch
10
+ import torch.nn.functional as F
11
+
12
+ from scipy.signal import kaiser
13
+
14
+
15
+ def design_prototype_filter(taps=62, cutoff_ratio=0.15, beta=9.0):
16
+ """Design prototype filter for PQMF.
17
+ This method is based on `A Kaiser window approach for the design of prototype
18
+ filters of cosine modulated filterbanks`_.
19
+ Args:
20
+ taps (int): The number of filter taps.
21
+ cutoff_ratio (float): Cut-off frequency ratio.
22
+ beta (float): Beta coefficient for kaiser window.
23
+ Returns:
24
+ ndarray: Impluse response of prototype filter (taps + 1,).
25
+ .. _`A Kaiser window approach for the design of prototype filters of cosine modulated filterbanks`:
26
+ https://ieeexplore.ieee.org/abstract/document/681427
27
+ """
28
+ # check the arguments are valid
29
+ assert taps % 2 == 0, "The number of taps mush be even number."
30
+ assert 0.0 < cutoff_ratio < 1.0, "Cutoff ratio must be > 0.0 and < 1.0."
31
+
32
+ # make initial filter
33
+ omega_c = np.pi * cutoff_ratio
34
+ with np.errstate(invalid='ignore'):
35
+ h_i = np.sin(omega_c * (np.arange(taps + 1) - 0.5 * taps)) \
36
+ / (np.pi * (np.arange(taps + 1) - 0.5 * taps))
37
+ h_i[taps // 2] = np.cos(0) * cutoff_ratio # fix nan due to indeterminate form
38
+
39
+ # apply kaiser window
40
+ w = kaiser(taps + 1, beta)
41
+ h = h_i * w
42
+
43
+ return h
44
+
45
+
46
+ class PQMF(torch.nn.Module):
47
+ """PQMF module.
48
+ This module is based on `Near-perfect-reconstruction pseudo-QMF banks`_.
49
+ .. _`Near-perfect-reconstruction pseudo-QMF banks`:
50
+ https://ieeexplore.ieee.org/document/258122
51
+ """
52
+
53
+ def __init__(self, device, subbands=4, taps=62, cutoff_ratio=0.15, beta=9.0):
54
+ """Initilize PQMF module.
55
+ Args:
56
+ subbands (int): The number of subbands.
57
+ taps (int): The number of filter taps.
58
+ cutoff_ratio (float): Cut-off frequency ratio.
59
+ beta (float): Beta coefficient for kaiser window.
60
+ """
61
+ super(PQMF, self).__init__()
62
+
63
+ # define filter coefficient
64
+ h_proto = design_prototype_filter(taps, cutoff_ratio, beta)
65
+ h_analysis = np.zeros((subbands, len(h_proto)))
66
+ h_synthesis = np.zeros((subbands, len(h_proto)))
67
+ for k in range(subbands):
68
+ h_analysis[k] = 2 * h_proto * np.cos(
69
+ (2 * k + 1) * (np.pi / (2 * subbands)) *
70
+ (np.arange(taps + 1) - ((taps - 1) / 2)) +
71
+ (-1) ** k * np.pi / 4)
72
+ h_synthesis[k] = 2 * h_proto * np.cos(
73
+ (2 * k + 1) * (np.pi / (2 * subbands)) *
74
+ (np.arange(taps + 1) - ((taps - 1) / 2)) -
75
+ (-1) ** k * np.pi / 4)
76
+
77
+ # convert to tensor
78
+ analysis_filter = torch.from_numpy(h_analysis).float().unsqueeze(1).to(device)
79
+ synthesis_filter = torch.from_numpy(h_synthesis).float().unsqueeze(0).to(device)
80
+
81
+ # register coefficients as beffer
82
+ self.register_buffer("analysis_filter", analysis_filter)
83
+ self.register_buffer("synthesis_filter", synthesis_filter)
84
+
85
+ # filter for downsampling & upsampling
86
+ updown_filter = torch.zeros((subbands, subbands, subbands)).float().to(device)
87
+ for k in range(subbands):
88
+ updown_filter[k, k, 0] = 1.0
89
+ self.register_buffer("updown_filter", updown_filter)
90
+ self.subbands = subbands
91
+
92
+ # keep padding info
93
+ self.pad_fn = torch.nn.ConstantPad1d(taps // 2, 0.0)
94
+
95
+ def analysis(self, x):
96
+ """Analysis with PQMF.
97
+ Args:
98
+ x (Tensor): Input tensor (B, 1, T).
99
+ Returns:
100
+ Tensor: Output tensor (B, subbands, T // subbands).
101
+ """
102
+ x = F.conv1d(self.pad_fn(x), self.analysis_filter)
103
+ return F.conv1d(x, self.updown_filter, stride=self.subbands)
104
+
105
+ def synthesis(self, x):
106
+ """Synthesis with PQMF.
107
+ Args:
108
+ x (Tensor): Input tensor (B, subbands, T // subbands).
109
+ Returns:
110
+ Tensor: Output tensor (B, 1, T).
111
+ """
112
+ # NOTE(kan-bayashi): Power will be dreased so here multipy by # subbands.
113
+ # Not sure this is the correct way, it is better to check again.
114
+ # TODO(kan-bayashi): Understand the reconstruction procedure
115
+ x = F.conv_transpose1d(x, self.updown_filter * self.subbands, stride=self.subbands)
116
+ return F.conv1d(self.pad_fn(x), self.synthesis_filter)
preprocess.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import text
3
+ from utils import load_filepaths_and_text
4
+
5
+ if __name__ == '__main__':
6
+ parser = argparse.ArgumentParser()
7
+ parser.add_argument("--out_extension", default="cleaned")
8
+ parser.add_argument("--text_index", default=1, type=int)
9
+ parser.add_argument("--filelists", nargs="+", default=["filelists/ljs_audio_text_val_filelist.txt", "filelists/ljs_audio_text_test_filelist.txt"])
10
+ parser.add_argument("--text_cleaners", nargs="+", default=["english_cleaners2"])
11
+
12
+ args = parser.parse_args()
13
+
14
+
15
+ for filelist in args.filelists:
16
+ print("START:", filelist)
17
+ filepaths_and_text = load_filepaths_and_text(filelist)
18
+ for i in range(len(filepaths_and_text)):
19
+ original_text = filepaths_and_text[i][args.text_index]
20
+ cleaned_text = text._clean_text(original_text, args.text_cleaners)
21
+ filepaths_and_text[i][args.text_index] = cleaned_text
22
+
23
+ new_filelist = filelist + "." + args.out_extension
24
+ with open(new_filelist, "w", encoding="utf-8") as f:
25
+ f.writelines(["|".join(x) + "\n" for x in filepaths_and_text])
requirements.txt ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Cython==0.29.21
2
+ librosa==0.8.0
3
+ matplotlib==3.3.1
4
+ numpy==1.18.5
5
+ phonemizer==2.2.1
6
+ scipy==1.5.2
7
+ torch
8
+ torchvision
9
+ Unidecode==1.1.1
10
+ pypinyin
11
+ jieba
12
+ cn2an
13
+ eng_to_ipa
14
+ inflect
15
+ langid
save_model/config.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 10,
4
+ "eval_interval": 100,
5
+ "seed": 1234,
6
+ "epochs": 1500,
7
+ "learning_rate": 2e-4,
8
+ "betas": [0.8, 0.99],
9
+ "eps": 1e-9,
10
+ "batch_size": 50,
11
+ "fp16_run": false,
12
+ "lr_decay": 0.999875,
13
+ "segment_size": 8192,
14
+ "init_lr_ratio": 1,
15
+ "warmup_epochs": 0,
16
+ "c_mel": 45,
17
+ "c_kl": 1.0,
18
+ "fft_sizes": [384, 683, 171],
19
+ "hop_sizes": [30, 60, 10],
20
+ "win_lengths": [150, 300, 60],
21
+ "window": "hann_window"
22
+ },
23
+ "data": {
24
+ "training_files":"filelists/genshin_cn_en.txt.cleaned",
25
+ "validation_files":"filelists/genshin_cn_en.txt.cleaned",
26
+ "text_cleaners":["zh_en_cleaners"],
27
+ "max_wav_value": 32768.0,
28
+ "sampling_rate": 22050,
29
+ "filter_length": 1024,
30
+ "hop_length": 256,
31
+ "win_length": 1024,
32
+ "n_mel_channels": 80,
33
+ "mel_fmin": 0.0,
34
+ "mel_fmax": null,
35
+ "add_blank": true,
36
+ "n_speakers": 50,
37
+ "cleaned_text": true
38
+ },
39
+ "model": {
40
+ "ms_istft_vits": true,
41
+ "mb_istft_vits": false,
42
+ "istft_vits": false,
43
+ "subbands": 4,
44
+ "gen_istft_n_fft": 16,
45
+ "gen_istft_hop_size": 4,
46
+ "inter_channels": 192,
47
+ "hidden_channels": 192,
48
+ "filter_channels": 768,
49
+ "n_heads": 2,
50
+ "n_layers": 6,
51
+ "kernel_size": 3,
52
+ "p_dropout": 0.1,
53
+ "resblock": "1",
54
+ "resblock_kernel_sizes": [3,7,11],
55
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
56
+ "upsample_rates": [4,4],
57
+ "upsample_initial_channel": 512,
58
+ "upsample_kernel_sizes": [16,16],
59
+ "n_layers_q": 3,
60
+ "use_spectral_norm": false,
61
+ "gin_channels": 256,
62
+ "use_sdp": false
63
+ }
64
+
65
+ }
66
+
save_model/model.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:30a5615c6a6d25adac34a8708f7cf417c411044bffd213e3ecc62106505f2c2f
3
+ size 456059188
save_model/npclists.txt ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 0|丽莎
2
+ 1|派蒙
3
+ 2|香菱
4
+ 3|神里绫华
5
+ 4|北斗
6
+ 5|枫原万叶
7
+ 6|莫娜
8
+ 7|刻晴
9
+ 8|七七
10
+ 9|神里绫人
11
+ 10|可莉
12
+ 11|诺艾尔
13
+ 12|甘雨
14
+ 13|凝光
15
+ 14|荒泷一斗
16
+ 15|烟绯
17
+ 16|夜兰
18
+ 17|凯亚
19
+ 18|钟离
20
+ 19|温迪
21
+ 20|琴
22
+ 21|芭芭拉
23
+ 22|罗莎莉亚
24
+ 23|菲谢尔
25
+ 24|辛焱
26
+ 25|阿贝多
27
+ 26|珊瑚宫心海
28
+ 27|宵宫
29
+ 28|班尼特
30
+ 29|雷泽
31
+ 30|鹿野院平藏
32
+ 31|优菈
33
+ 32|安柏
34
+ 33|早柚
35
+ 34|重云
36
+ 35|申鹤
37
+ 36|云堇
38
+ 37|行秋
39
+ 38|魈
40
+ 39|胡桃
41
+ 40|托马
42
+ 41|五郎
43
+ 42|八重神子
44
+ 43|九条裟罗
45
+ 44|达达利亚
46
+ 45|迪卢克
47
+ 46|迪奥娜
48
+ 47|砂糖
49
+ 48|雷电将军
stft.py ADDED
@@ -0,0 +1,209 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ BSD 3-Clause License
3
+ Copyright (c) 2017, Prem Seetharaman
4
+ All rights reserved.
5
+ * Redistribution and use in source and binary forms, with or without
6
+ modification, are permitted provided that the following conditions are met:
7
+ * Redistributions of source code must retain the above copyright notice,
8
+ this list of conditions and the following disclaimer.
9
+ * Redistributions in binary form must reproduce the above copyright notice, this
10
+ list of conditions and the following disclaimer in the
11
+ documentation and/or other materials provided with the distribution.
12
+ * Neither the name of the copyright holder nor the names of its
13
+ contributors may be used to endorse or promote products derived from this
14
+ software without specific prior written permission.
15
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
16
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
17
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
18
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
19
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
20
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
21
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
22
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
24
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25
+ """
26
+
27
+ import torch
28
+ import numpy as np
29
+ import torch.nn.functional as F
30
+ from torch.autograd import Variable
31
+ from scipy.signal import get_window
32
+ from librosa.util import pad_center, tiny
33
+ import librosa.util as librosa_util
34
+
35
+ def window_sumsquare(window, n_frames, hop_length=200, win_length=800,
36
+ n_fft=800, dtype=np.float32, norm=None):
37
+ """
38
+ # from librosa 0.6
39
+ Compute the sum-square envelope of a window function at a given hop length.
40
+ This is used to estimate modulation effects induced by windowing
41
+ observations in short-time fourier transforms.
42
+ Parameters
43
+ ----------
44
+ window : string, tuple, number, callable, or list-like
45
+ Window specification, as in `get_window`
46
+ n_frames : int > 0
47
+ The number of analysis frames
48
+ hop_length : int > 0
49
+ The number of samples to advance between frames
50
+ win_length : [optional]
51
+ The length of the window function. By default, this matches `n_fft`.
52
+ n_fft : int > 0
53
+ The length of each analysis frame.
54
+ dtype : np.dtype
55
+ The data type of the output
56
+ Returns
57
+ -------
58
+ wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
59
+ The sum-squared envelope of the window function
60
+ """
61
+ if win_length is None:
62
+ win_length = n_fft
63
+
64
+ n = n_fft + hop_length * (n_frames - 1)
65
+ x = np.zeros(n, dtype=dtype)
66
+
67
+ # Compute the squared window at the desired length
68
+ win_sq = get_window(window, win_length, fftbins=True)
69
+ win_sq = librosa_util.normalize(win_sq, norm=norm)**2
70
+ win_sq = librosa_util.pad_center(win_sq, n_fft)
71
+
72
+ # Fill the envelope
73
+ for i in range(n_frames):
74
+ sample = i * hop_length
75
+ x[sample:min(n, sample + n_fft)] += win_sq[:max(0, min(n_fft, n - sample))]
76
+ return x
77
+
78
+
79
+ class STFT(torch.nn.Module):
80
+ """adapted from Prem Seetharaman's https://github.com/pseeth/pytorch-stft"""
81
+ def __init__(self, filter_length=800, hop_length=200, win_length=800,
82
+ window='hann'):
83
+ super(STFT, self).__init__()
84
+ self.filter_length = filter_length
85
+ self.hop_length = hop_length
86
+ self.win_length = win_length
87
+ self.window = window
88
+ self.forward_transform = None
89
+ scale = self.filter_length / self.hop_length
90
+ fourier_basis = np.fft.fft(np.eye(self.filter_length))
91
+
92
+ cutoff = int((self.filter_length / 2 + 1))
93
+ fourier_basis = np.vstack([np.real(fourier_basis[:cutoff, :]),
94
+ np.imag(fourier_basis[:cutoff, :])])
95
+
96
+ forward_basis = torch.FloatTensor(fourier_basis[:, None, :])
97
+ inverse_basis = torch.FloatTensor(
98
+ np.linalg.pinv(scale * fourier_basis).T[:, None, :])
99
+
100
+ if window is not None:
101
+ assert(filter_length >= win_length)
102
+ # get window and zero center pad it to filter_length
103
+ fft_window = get_window(window, win_length, fftbins=True)
104
+ fft_window = pad_center(fft_window, filter_length)
105
+ fft_window = torch.from_numpy(fft_window).float()
106
+
107
+ # window the bases
108
+ forward_basis *= fft_window
109
+ inverse_basis *= fft_window
110
+
111
+ self.register_buffer('forward_basis', forward_basis.float())
112
+ self.register_buffer('inverse_basis', inverse_basis.float())
113
+
114
+ def transform(self, input_data):
115
+ num_batches = input_data.size(0)
116
+ num_samples = input_data.size(1)
117
+
118
+ self.num_samples = num_samples
119
+
120
+ # similar to librosa, reflect-pad the input
121
+ input_data = input_data.view(num_batches, 1, num_samples)
122
+ input_data = F.pad(
123
+ input_data.unsqueeze(1),
124
+ (int(self.filter_length / 2), int(self.filter_length / 2), 0, 0),
125
+ mode='reflect')
126
+ input_data = input_data.squeeze(1)
127
+
128
+ forward_transform = F.conv1d(
129
+ input_data,
130
+ Variable(self.forward_basis, requires_grad=False),
131
+ stride=self.hop_length,
132
+ padding=0)
133
+
134
+ cutoff = int((self.filter_length / 2) + 1)
135
+ real_part = forward_transform[:, :cutoff, :]
136
+ imag_part = forward_transform[:, cutoff:, :]
137
+
138
+ magnitude = torch.sqrt(real_part**2 + imag_part**2)
139
+ phase = torch.autograd.Variable(
140
+ torch.atan2(imag_part.data, real_part.data))
141
+
142
+ return magnitude, phase
143
+
144
+ def inverse(self, magnitude, phase):
145
+ recombine_magnitude_phase = torch.cat(
146
+ [magnitude*torch.cos(phase), magnitude*torch.sin(phase)], dim=1)
147
+
148
+ inverse_transform = F.conv_transpose1d(
149
+ recombine_magnitude_phase,
150
+ Variable(self.inverse_basis, requires_grad=False),
151
+ stride=self.hop_length,
152
+ padding=0)
153
+
154
+ if self.window is not None:
155
+ window_sum = window_sumsquare(
156
+ self.window, magnitude.size(-1), hop_length=self.hop_length,
157
+ win_length=self.win_length, n_fft=self.filter_length,
158
+ dtype=np.float32)
159
+ # remove modulation effects
160
+ approx_nonzero_indices = torch.from_numpy(
161
+ np.where(window_sum > tiny(window_sum))[0])
162
+ window_sum = torch.autograd.Variable(
163
+ torch.from_numpy(window_sum), requires_grad=False)
164
+ window_sum = window_sum.to(inverse_transform.device()) if magnitude.is_cuda else window_sum
165
+ inverse_transform[:, :, approx_nonzero_indices] /= window_sum[approx_nonzero_indices]
166
+
167
+ # scale by hop ratio
168
+ inverse_transform *= float(self.filter_length) / self.hop_length
169
+
170
+ inverse_transform = inverse_transform[:, :, int(self.filter_length/2):]
171
+ inverse_transform = inverse_transform[:, :, :-int(self.filter_length/2):]
172
+
173
+ return inverse_transform
174
+
175
+ def forward(self, input_data):
176
+ self.magnitude, self.phase = self.transform(input_data)
177
+ reconstruction = self.inverse(self.magnitude, self.phase)
178
+ return reconstruction
179
+
180
+
181
+ class TorchSTFT(torch.nn.Module):
182
+ def __init__(self, filter_length=800, hop_length=200, win_length=800, window='hann'):
183
+ super().__init__()
184
+ self.filter_length = filter_length
185
+ self.hop_length = hop_length
186
+ self.win_length = win_length
187
+ self.window = torch.from_numpy(get_window(window, win_length, fftbins=True).astype(np.float32))
188
+
189
+ def transform(self, input_data):
190
+ forward_transform = torch.stft(
191
+ input_data,
192
+ self.filter_length, self.hop_length, self.win_length, window=self.window,
193
+ return_complex=True)
194
+
195
+ return torch.abs(forward_transform), torch.angle(forward_transform)
196
+
197
+ def inverse(self, magnitude, phase):
198
+ inverse_transform = torch.istft(
199
+ magnitude * torch.exp(phase * 1j),
200
+ self.filter_length, self.hop_length, self.win_length, window=self.window.to(magnitude.device))
201
+
202
+ return inverse_transform.unsqueeze(-2) # unsqueeze to stay consistent with conv_transpose1d implementation
203
+
204
+ def forward(self, input_data):
205
+ self.magnitude, self.phase = self.transform(input_data)
206
+ reconstruction = self.inverse(self.magnitude, self.phase)
207
+ return reconstruction
208
+
209
+
stft_loss.py ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+
3
+ # Copyright 2019 Tomoki Hayashi
4
+ # MIT License (https://opensource.org/licenses/MIT)
5
+
6
+ """STFT-based Loss modules."""
7
+
8
+ import torch
9
+ import torch.nn.functional as F
10
+
11
+
12
+ def stft(x, fft_size, hop_size, win_length, window):
13
+ """Perform STFT and convert to magnitude spectrogram.
14
+ Args:
15
+ x (Tensor): Input signal tensor (B, T).
16
+ fft_size (int): FFT size.
17
+ hop_size (int): Hop size.
18
+ win_length (int): Window length.
19
+ window (str): Window function type.
20
+ Returns:
21
+ Tensor: Magnitude spectrogram (B, #frames, fft_size // 2 + 1).
22
+ """
23
+ x_stft = torch.stft(x, fft_size, hop_size, win_length, window.to(x.device))
24
+ real = x_stft[..., 0]
25
+ imag = x_stft[..., 1]
26
+
27
+ # NOTE(kan-bayashi): clamp is needed to avoid nan or inf
28
+ return torch.sqrt(torch.clamp(real ** 2 + imag ** 2, min=1e-7)).transpose(2, 1)
29
+
30
+
31
+ class SpectralConvergengeLoss(torch.nn.Module):
32
+ """Spectral convergence loss module."""
33
+
34
+ def __init__(self):
35
+ """Initilize spectral convergence loss module."""
36
+ super(SpectralConvergengeLoss, self).__init__()
37
+
38
+ def forward(self, x_mag, y_mag):
39
+ """Calculate forward propagation.
40
+ Args:
41
+ x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins).
42
+ y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins).
43
+ Returns:
44
+ Tensor: Spectral convergence loss value.
45
+ """
46
+ return torch.norm(y_mag - x_mag, p="fro") / torch.norm(y_mag, p="fro")
47
+
48
+
49
+ class LogSTFTMagnitudeLoss(torch.nn.Module):
50
+ """Log STFT magnitude loss module."""
51
+
52
+ def __init__(self):
53
+ """Initilize los STFT magnitude loss module."""
54
+ super(LogSTFTMagnitudeLoss, self).__init__()
55
+
56
+ def forward(self, x_mag, y_mag):
57
+ """Calculate forward propagation.
58
+ Args:
59
+ x_mag (Tensor): Magnitude spectrogram of predicted signal (B, #frames, #freq_bins).
60
+ y_mag (Tensor): Magnitude spectrogram of groundtruth signal (B, #frames, #freq_bins).
61
+ Returns:
62
+ Tensor: Log STFT magnitude loss value.
63
+ """
64
+ return F.l1_loss(torch.log(y_mag), torch.log(x_mag))
65
+
66
+
67
+ class STFTLoss(torch.nn.Module):
68
+ """STFT loss module."""
69
+
70
+ def __init__(self, fft_size=1024, shift_size=120, win_length=600, window="hann_window"):
71
+ """Initialize STFT loss module."""
72
+ super(STFTLoss, self).__init__()
73
+ self.fft_size = fft_size
74
+ self.shift_size = shift_size
75
+ self.win_length = win_length
76
+ self.window = getattr(torch, window)(win_length)
77
+ self.spectral_convergenge_loss = SpectralConvergengeLoss()
78
+ self.log_stft_magnitude_loss = LogSTFTMagnitudeLoss()
79
+
80
+ def forward(self, x, y):
81
+ """Calculate forward propagation.
82
+ Args:
83
+ x (Tensor): Predicted signal (B, T).
84
+ y (Tensor): Groundtruth signal (B, T).
85
+ Returns:
86
+ Tensor: Spectral convergence loss value.
87
+ Tensor: Log STFT magnitude loss value.
88
+ """
89
+ x_mag = stft(x, self.fft_size, self.shift_size, self.win_length, self.window)
90
+ y_mag = stft(y, self.fft_size, self.shift_size, self.win_length, self.window)
91
+ sc_loss = self.spectral_convergenge_loss(x_mag, y_mag)
92
+ mag_loss = self.log_stft_magnitude_loss(x_mag, y_mag)
93
+
94
+ return sc_loss, mag_loss
95
+
96
+
97
+ class MultiResolutionSTFTLoss(torch.nn.Module):
98
+ """Multi resolution STFT loss module."""
99
+
100
+ def __init__(self,
101
+ fft_sizes=[1024, 2048, 512],
102
+ hop_sizes=[120, 240, 50],
103
+ win_lengths=[600, 1200, 240],
104
+ window="hann_window"):
105
+ """Initialize Multi resolution STFT loss module.
106
+ Args:
107
+ fft_sizes (list): List of FFT sizes.
108
+ hop_sizes (list): List of hop sizes.
109
+ win_lengths (list): List of window lengths.
110
+ window (str): Window function type.
111
+ """
112
+ super(MultiResolutionSTFTLoss, self).__init__()
113
+ assert len(fft_sizes) == len(hop_sizes) == len(win_lengths)
114
+ self.stft_losses = torch.nn.ModuleList()
115
+ for fs, ss, wl in zip(fft_sizes, hop_sizes, win_lengths):
116
+ self.stft_losses += [STFTLoss(fs, ss, wl, window)]
117
+
118
+ def forward(self, x, y):
119
+ """Calculate forward propagation.
120
+ Args:
121
+ x (Tensor): Predicted signal (B, T).
122
+ y (Tensor): Groundtruth signal (B, T).
123
+ Returns:
124
+ Tensor: Multi resolution spectral convergence loss value.
125
+ Tensor: Multi resolution log STFT magnitude loss value.
126
+ """
127
+ sc_loss = 0.0
128
+ mag_loss = 0.0
129
+ for f in self.stft_losses:
130
+ sc_l, mag_l = f(x, y)
131
+ sc_loss += sc_l
132
+ mag_loss += mag_l
133
+ sc_loss /= len(self.stft_losses)
134
+ mag_loss /= len(self.stft_losses)
135
+
136
+ return sc_loss, mag_loss
text/LICENSE ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright (c) 2017 Keith Ito
2
+
3
+ Permission is hereby granted, free of charge, to any person obtaining a copy
4
+ of this software and associated documentation files (the "Software"), to deal
5
+ in the Software without restriction, including without limitation the rights
6
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7
+ copies of the Software, and to permit persons to whom the Software is
8
+ furnished to do so, subject to the following conditions:
9
+
10
+ The above copyright notice and this permission notice shall be included in
11
+ all copies or substantial portions of the Software.
12
+
13
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
19
+ THE SOFTWARE.
text/__init__.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+ from text import cleaners
3
+ from text.symbols import symbols
4
+
5
+
6
+ # Mappings from symbol to numeric ID and vice versa:
7
+ _symbol_to_id = {s: i for i, s in enumerate(symbols)}
8
+ _id_to_symbol = {i: s for i, s in enumerate(symbols)}
9
+
10
+
11
+ def text_to_sequence(text, cleaner_names):
12
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
13
+ Args:
14
+ text: string to convert to a sequence
15
+ cleaner_names: names of the cleaner functions to run the text through
16
+ Returns:
17
+ List of integers corresponding to the symbols in the text
18
+ '''
19
+ sequence = []
20
+
21
+ clean_text = _clean_text(text, cleaner_names)
22
+ for symbol in clean_text:
23
+ if symbol not in _symbol_to_id.keys():
24
+ continue
25
+ symbol_id = _symbol_to_id[symbol]
26
+ sequence += [symbol_id]
27
+ return sequence
28
+
29
+
30
+ def cleaned_text_to_sequence(cleaned_text):
31
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
32
+ Args:
33
+ text: string to convert to a sequence
34
+ Returns:
35
+ List of integers corresponding to the symbols in the text
36
+ '''
37
+ sequence = [_symbol_to_id[symbol] for symbol in cleaned_text if symbol in _symbol_to_id.keys()]
38
+ return sequence
39
+
40
+
41
+ def sequence_to_text(sequence):
42
+ '''Converts a sequence of IDs back to a string'''
43
+ result = ''
44
+ for symbol_id in sequence:
45
+ s = _id_to_symbol[symbol_id]
46
+ result += s
47
+ return result
48
+
49
+
50
+ def _clean_text(text, cleaner_names):
51
+ for name in cleaner_names:
52
+ cleaner = getattr(cleaners, name)
53
+ if not cleaner:
54
+ raise Exception('Unknown cleaner: %s' % name)
55
+ text = cleaner(text)
56
+ return text
text/cleaners.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ #from text.japanese import japanese_to_romaji_with_accent, japanese_to_ipa, japanese_to_ipa2, japanese_to_ipa3
3
+ from text.mandarin import number_to_chinese, chinese_to_bopomofo, latin_to_bopomofo, chinese_to_romaji, chinese_to_lazy_ipa, chinese_to_ipa, chinese_to_ipa2
4
+ # from text.sanskrit import devanagari_to_ipa
5
+ from text.english import english_to_lazy_ipa, english_to_ipa2
6
+ # from text.thai import num_to_thai, latin_to_thai
7
+ # from text.shanghainese import shanghainese_to_ipa
8
+ # from text.cantonese import cantonese_to_ipa
9
+ # from text.ngu_dialect import ngu_dialect_to_ipa
10
+
11
+
12
+ def japanese_cleaners(text):
13
+ text = japanese_to_romaji_with_accent(text)
14
+ if re.match('[A-Za-z]', text[-1]):
15
+ text += '.'
16
+ return text
17
+
18
+
19
+ def japanese_cleaners2(text):
20
+ return japanese_cleaners(text).replace('ts', 'ʦ').replace('...', '…')
21
+
22
+
23
+ def korean_cleaners(text):
24
+ '''Pipeline for Korean text'''
25
+ text = latin_to_hangul(text)
26
+ text = number_to_hangul(text)
27
+ text = divide_hangul(text)
28
+ if re.match('[\u3131-\u3163]', text[-1]):
29
+ text += '.'
30
+ return text
31
+
32
+
33
+ def chinese_cleaners(text):
34
+ '''Pipeline for Chinese text'''
35
+ text = number_to_chinese(text)
36
+ text = chinese_to_bopomofo(text)
37
+ text = latin_to_bopomofo(text)
38
+ if re.match('[ˉˊˇˋ˙]', text[-1]):
39
+ text += '。'
40
+ return text
41
+
42
+
43
+ def zh_ja_mixture_cleaners(text):
44
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
45
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
46
+ for chinese_text in chinese_texts:
47
+ cleaned_text = chinese_to_romaji(chinese_text[4:-4])
48
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
49
+ for japanese_text in japanese_texts:
50
+ cleaned_text = japanese_to_romaji_with_accent(
51
+ japanese_text[4:-4]).replace('ts', 'ʦ').replace('u', 'ɯ').replace('...', '…')
52
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
53
+ text = text[:-1]
54
+ if re.match('[A-Za-zɯɹəɥ→↓↑]', text[-1]):
55
+ text += '.'
56
+ return text
57
+
58
+
59
+ def sanskrit_cleaners(text):
60
+ text = text.replace('॥', '।').replace('ॐ', 'ओम्')
61
+ if text[-1] != '।':
62
+ text += ' ।'
63
+ return text
64
+
65
+ def zh_en_cleaners(text):
66
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
67
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
68
+ for chinese_text in chinese_texts:
69
+ cleaned_text = chinese_to_lazy_ipa(chinese_text[4:-4])
70
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
71
+ for english_text in english_texts:
72
+ cleaned_text = english_to_lazy_ipa(english_text[4:-4])
73
+ text = text.replace(english_text, cleaned_text+' ', 1)
74
+ text = text[:-1]
75
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
76
+ text += '.'
77
+ return text
78
+
79
+ def cjks_cleaners(text):
80
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
81
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
82
+ korean_texts = re.findall(r'\[KO\].*?\[KO\]', text)
83
+ sanskrit_texts = re.findall(r'\[SA\].*?\[SA\]', text)
84
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
85
+ for chinese_text in chinese_texts:
86
+ cleaned_text = chinese_to_lazy_ipa(chinese_text[4:-4])
87
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
88
+ for japanese_text in japanese_texts:
89
+ cleaned_text = japanese_to_ipa(japanese_text[4:-4])
90
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
91
+ for korean_text in korean_texts:
92
+ cleaned_text = korean_to_lazy_ipa(korean_text[4:-4])
93
+ text = text.replace(korean_text, cleaned_text+' ', 1)
94
+ for sanskrit_text in sanskrit_texts:
95
+ cleaned_text = devanagari_to_ipa(sanskrit_text[4:-4])
96
+ text = text.replace(sanskrit_text, cleaned_text+' ', 1)
97
+ for english_text in english_texts:
98
+ cleaned_text = english_to_lazy_ipa(english_text[4:-4])
99
+ text = text.replace(english_text, cleaned_text+' ', 1)
100
+ text = text[:-1]
101
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
102
+ text += '.'
103
+ return text
104
+
105
+
106
+ def cjke_cleaners(text):
107
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
108
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
109
+ korean_texts = re.findall(r'\[KO\].*?\[KO\]', text)
110
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
111
+ for chinese_text in chinese_texts:
112
+ cleaned_text = chinese_to_lazy_ipa(chinese_text[4:-4])
113
+ cleaned_text = cleaned_text.replace(
114
+ 'ʧ', 'tʃ').replace('ʦ', 'ts').replace('ɥan', 'ɥæn')
115
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
116
+ for japanese_text in japanese_texts:
117
+ cleaned_text = japanese_to_ipa(japanese_text[4:-4])
118
+ cleaned_text = cleaned_text.replace('ʧ', 'tʃ').replace(
119
+ 'ʦ', 'ts').replace('ɥan', 'ɥæn').replace('ʥ', 'dz')
120
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
121
+ for korean_text in korean_texts:
122
+ cleaned_text = korean_to_ipa(korean_text[4:-4])
123
+ text = text.replace(korean_text, cleaned_text+' ', 1)
124
+ for english_text in english_texts:
125
+ cleaned_text = english_to_ipa2(english_text[4:-4])
126
+ cleaned_text = cleaned_text.replace('ɑ', 'a').replace(
127
+ 'ɔ', 'o').replace('ɛ', 'e').replace('ɪ', 'i').replace('ʊ', 'u')
128
+ text = text.replace(english_text, cleaned_text+' ', 1)
129
+ text = text[:-1]
130
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
131
+ text += '.'
132
+ return text
133
+
134
+
135
+ def cjke_cleaners2(text):
136
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
137
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
138
+ korean_texts = re.findall(r'\[KO\].*?\[KO\]', text)
139
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
140
+ for chinese_text in chinese_texts:
141
+ cleaned_text = chinese_to_ipa(chinese_text[4:-4])
142
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
143
+ for japanese_text in japanese_texts:
144
+ cleaned_text = japanese_to_ipa2(japanese_text[4:-4])
145
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
146
+ for korean_text in korean_texts:
147
+ cleaned_text = korean_to_ipa(korean_text[4:-4])
148
+ text = text.replace(korean_text, cleaned_text+' ', 1)
149
+ for english_text in english_texts:
150
+ cleaned_text = english_to_ipa2(english_text[4:-4])
151
+ text = text.replace(english_text, cleaned_text+' ', 1)
152
+ text = text[:-1]
153
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
154
+ text += '.'
155
+ return text
156
+
157
+
158
+ def thai_cleaners(text):
159
+ text = num_to_thai(text)
160
+ text = latin_to_thai(text)
161
+ return text
162
+
163
+
164
+ def shanghainese_cleaners(text):
165
+ text = shanghainese_to_ipa(text)
166
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
167
+ text += '.'
168
+ return text
169
+
170
+
171
+ def chinese_dialect_cleaners(text):
172
+ text = re.sub(r'\[MD\](.*?)\[MD\]',
173
+ lambda x: chinese_to_ipa2(x.group(1))+' ', text)
174
+ text = re.sub(r'\[TW\](.*?)\[TW\]',
175
+ lambda x: chinese_to_ipa2(x.group(1), True)+' ', text)
176
+ text = re.sub(r'\[JA\](.*?)\[JA\]',
177
+ lambda x: japanese_to_ipa3(x.group(1)).replace('Q', 'ʔ')+' ', text)
178
+ text = re.sub(r'\[SH\](.*?)\[SH\]', lambda x: shanghainese_to_ipa(x.group(1)).replace('1', '˥˧').replace('5',
179
+ '˧˧˦').replace('6', '˩˩˧').replace('7', '˥').replace('8', '˩˨').replace('ᴀ', 'ɐ').replace('ᴇ', 'e')+' ', text)
180
+ text = re.sub(r'\[GD\](.*?)\[GD\]',
181
+ lambda x: cantonese_to_ipa(x.group(1))+' ', text)
182
+ text = re.sub(r'\[EN\](.*?)\[EN\]',
183
+ lambda x: english_to_lazy_ipa2(x.group(1))+' ', text)
184
+ text = re.sub(r'\[([A-Z]{2})\](.*?)\[\1\]', lambda x: ngu_dialect_to_ipa(x.group(2), x.group(
185
+ 1)).replace('ʣ', 'dz').replace('ʥ', 'dʑ').replace('ʦ', 'ts').replace('ʨ', 'tɕ')+' ', text)
186
+ text = re.sub(r'\s+$', '', text)
187
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
188
+ return text
text/english.py ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+
3
+ '''
4
+ Cleaners are transformations that run over the input text at both training and eval time.
5
+
6
+ Cleaners can be selected by passing a comma-delimited list of cleaner names as the "cleaners"
7
+ hyperparameter. Some cleaners are English-specific. You'll typically want to use:
8
+ 1. "english_cleaners" for English text
9
+ 2. "transliteration_cleaners" for non-English text that can be transliterated to ASCII using
10
+ the Unidecode library (https://pypi.python.org/pypi/Unidecode)
11
+ 3. "basic_cleaners" if you do not want to transliterate (in this case, you should also update
12
+ the symbols in symbols.py to match your data).
13
+ '''
14
+
15
+
16
+ # Regular expression matching whitespace:
17
+
18
+
19
+ import re
20
+ import inflect
21
+ from unidecode import unidecode
22
+ import eng_to_ipa as ipa
23
+ _inflect = inflect.engine()
24
+ _comma_number_re = re.compile(r'([0-9][0-9\,]+[0-9])')
25
+ _decimal_number_re = re.compile(r'([0-9]+\.[0-9]+)')
26
+ _pounds_re = re.compile(r'£([0-9\,]*[0-9]+)')
27
+ _dollars_re = re.compile(r'\$([0-9\.\,]*[0-9]+)')
28
+ _ordinal_re = re.compile(r'[0-9]+(st|nd|rd|th)')
29
+ _number_re = re.compile(r'[0-9]+')
30
+
31
+ # List of (regular expression, replacement) pairs for abbreviations:
32
+ _abbreviations = [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1]) for x in [
33
+ ('mrs', 'misess'),
34
+ ('mr', 'mister'),
35
+ ('dr', 'doctor'),
36
+ ('st', 'saint'),
37
+ ('co', 'company'),
38
+ ('jr', 'junior'),
39
+ ('maj', 'major'),
40
+ ('gen', 'general'),
41
+ ('drs', 'doctors'),
42
+ ('rev', 'reverend'),
43
+ ('lt', 'lieutenant'),
44
+ ('hon', 'honorable'),
45
+ ('sgt', 'sergeant'),
46
+ ('capt', 'captain'),
47
+ ('esq', 'esquire'),
48
+ ('ltd', 'limited'),
49
+ ('col', 'colonel'),
50
+ ('ft', 'fort'),
51
+ ]]
52
+
53
+
54
+ # List of (ipa, lazy ipa) pairs:
55
+ _lazy_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
56
+ ('r', 'ɹ'),
57
+ ('æ', 'e'),
58
+ ('ɑ', 'a'),
59
+ ('ɔ', 'o'),
60
+ ('ð', 'z'),
61
+ ('θ', 's'),
62
+ ('ɛ', 'e'),
63
+ ('ɪ', 'i'),
64
+ ('ʊ', 'u'),
65
+ ('ʒ', 'ʥ'),
66
+ ('ʤ', 'ʥ'),
67
+ ('ˈ', '↓'),
68
+ ]]
69
+
70
+ # List of (ipa, ipa2) pairs
71
+ _ipa_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
72
+ ('r', 'ɹ'),
73
+ ('ʤ', 'dʒ'),
74
+ ('ʧ', 'tʃ')
75
+ ]]
76
+
77
+
78
+ def expand_abbreviations(text):
79
+ for regex, replacement in _abbreviations:
80
+ text = re.sub(regex, replacement, text)
81
+ return text
82
+
83
+
84
+ def collapse_whitespace(text):
85
+ return re.sub(r'\s+', ' ', text)
86
+
87
+
88
+ def _remove_commas(m):
89
+ return m.group(1).replace(',', '')
90
+
91
+
92
+ def _expand_decimal_point(m):
93
+ return m.group(1).replace('.', ' point ')
94
+
95
+
96
+ def _expand_dollars(m):
97
+ match = m.group(1)
98
+ parts = match.split('.')
99
+ if len(parts) > 2:
100
+ return match + ' dollars' # Unexpected format
101
+ dollars = int(parts[0]) if parts[0] else 0
102
+ cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
103
+ if dollars and cents:
104
+ dollar_unit = 'dollar' if dollars == 1 else 'dollars'
105
+ cent_unit = 'cent' if cents == 1 else 'cents'
106
+ return '%s %s, %s %s' % (dollars, dollar_unit, cents, cent_unit)
107
+ elif dollars:
108
+ dollar_unit = 'dollar' if dollars == 1 else 'dollars'
109
+ return '%s %s' % (dollars, dollar_unit)
110
+ elif cents:
111
+ cent_unit = 'cent' if cents == 1 else 'cents'
112
+ return '%s %s' % (cents, cent_unit)
113
+ else:
114
+ return 'zero dollars'
115
+
116
+
117
+ def _expand_ordinal(m):
118
+ return _inflect.number_to_words(m.group(0))
119
+
120
+
121
+ def _expand_number(m):
122
+ num = int(m.group(0))
123
+ if num > 1000 and num < 3000:
124
+ if num == 2000:
125
+ return 'two thousand'
126
+ elif num > 2000 and num < 2010:
127
+ return 'two thousand ' + _inflect.number_to_words(num % 100)
128
+ elif num % 100 == 0:
129
+ return _inflect.number_to_words(num // 100) + ' hundred'
130
+ else:
131
+ return _inflect.number_to_words(num, andword='', zero='oh', group=2).replace(', ', ' ')
132
+ else:
133
+ return _inflect.number_to_words(num, andword='')
134
+
135
+
136
+ def normalize_numbers(text):
137
+ text = re.sub(_comma_number_re, _remove_commas, text)
138
+ text = re.sub(_pounds_re, r'\1 pounds', text)
139
+ text = re.sub(_dollars_re, _expand_dollars, text)
140
+ text = re.sub(_decimal_number_re, _expand_decimal_point, text)
141
+ text = re.sub(_ordinal_re, _expand_ordinal, text)
142
+ text = re.sub(_number_re, _expand_number, text)
143
+ return text
144
+
145
+
146
+ def mark_dark_l(text):
147
+ return re.sub(r'l([^aeiouæɑɔəɛɪʊ ]*(?: |$))', lambda x: 'ɫ'+x.group(1), text)
148
+
149
+
150
+ def english_to_ipa(text):
151
+ text = unidecode(text).lower()
152
+ text = expand_abbreviations(text)
153
+ text = normalize_numbers(text)
154
+ phonemes = ipa.convert(text)
155
+ phonemes = collapse_whitespace(phonemes)
156
+ return phonemes
157
+
158
+
159
+ def english_to_lazy_ipa(text):
160
+ text = english_to_ipa(text)
161
+ for regex, replacement in _lazy_ipa:
162
+ text = re.sub(regex, replacement, text)
163
+ return text
164
+
165
+
166
+ def english_to_ipa2(text):
167
+ text = english_to_ipa(text)
168
+ text = mark_dark_l(text)
169
+ for regex, replacement in _ipa_to_ipa2:
170
+ text = re.sub(regex, replacement, text)
171
+ return text.replace('...', '…')
text/japanese.py ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from unidecode import unidecode
3
+ import pyopenjtalk
4
+
5
+
6
+ # Regular expression matching Japanese without punctuation marks:
7
+ _japanese_characters = re.compile(
8
+ r'[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
9
+
10
+ # Regular expression matching non-Japanese characters or punctuation marks:
11
+ _japanese_marks = re.compile(
12
+ r'[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
13
+
14
+ # List of (symbol, Japanese) pairs for marks:
15
+ _symbols_to_japanese = [(re.compile('%s' % x[0]), x[1]) for x in [
16
+ ('%', 'パーセント')
17
+ ]]
18
+
19
+ # List of (romaji, ipa) pairs for marks:
20
+ _romaji_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
21
+ ('ts', 'ʦ'),
22
+ ('u', 'ɯ'),
23
+ ('j', 'ʥ'),
24
+ ('y', 'j'),
25
+ ('ni', 'n^i'),
26
+ ('nj', 'n^'),
27
+ ('hi', 'çi'),
28
+ ('hj', 'ç'),
29
+ ('f', 'ɸ'),
30
+ ('I', 'i*'),
31
+ ('U', 'ɯ*'),
32
+ ('r', 'ɾ')
33
+ ]]
34
+
35
+ # List of (romaji, ipa2) pairs for marks:
36
+ _romaji_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
37
+ ('u', 'ɯ'),
38
+ ('ʧ', 'tʃ'),
39
+ ('j', 'dʑ'),
40
+ ('y', 'j'),
41
+ ('ni', 'n^i'),
42
+ ('nj', 'n^'),
43
+ ('hi', 'çi'),
44
+ ('hj', 'ç'),
45
+ ('f', 'ɸ'),
46
+ ('I', 'i*'),
47
+ ('U', 'ɯ*'),
48
+ ('r', 'ɾ')
49
+ ]]
50
+
51
+ # List of (consonant, sokuon) pairs:
52
+ _real_sokuon = [(re.compile('%s' % x[0]), x[1]) for x in [
53
+ (r'Q([↑↓]*[kg])', r'k#\1'),
54
+ (r'Q([↑↓]*[tdjʧ])', r't#\1'),
55
+ (r'Q([↑↓]*[sʃ])', r's\1'),
56
+ (r'Q([↑↓]*[pb])', r'p#\1')
57
+ ]]
58
+
59
+ # List of (consonant, hatsuon) pairs:
60
+ _real_hatsuon = [(re.compile('%s' % x[0]), x[1]) for x in [
61
+ (r'N([↑↓]*[pbm])', r'm\1'),
62
+ (r'N([↑↓]*[ʧʥj])', r'n^\1'),
63
+ (r'N([↑↓]*[tdn])', r'n\1'),
64
+ (r'N([↑↓]*[kg])', r'ŋ\1')
65
+ ]]
66
+
67
+
68
+ def symbols_to_japanese(text):
69
+ for regex, replacement in _symbols_to_japanese:
70
+ text = re.sub(regex, replacement, text)
71
+ return text
72
+
73
+
74
+ def japanese_to_romaji_with_accent(text):
75
+ '''Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html'''
76
+ text = symbols_to_japanese(text)
77
+ sentences = re.split(_japanese_marks, text)
78
+ marks = re.findall(_japanese_marks, text)
79
+ text = ''
80
+ for i, sentence in enumerate(sentences):
81
+ if re.match(_japanese_characters, sentence):
82
+ if text != '':
83
+ text += ' '
84
+ labels = pyopenjtalk.extract_fullcontext(sentence)
85
+ for n, label in enumerate(labels):
86
+ phoneme = re.search(r'\-([^\+]*)\+', label).group(1)
87
+ if phoneme not in ['sil', 'pau']:
88
+ text += phoneme.replace('ch', 'ʧ').replace('sh',
89
+ 'ʃ').replace('cl', 'Q')
90
+ else:
91
+ continue
92
+ # n_moras = int(re.search(r'/F:(\d+)_', label).group(1))
93
+ a1 = int(re.search(r"/A:(\-?[0-9]+)\+", label).group(1))
94
+ a2 = int(re.search(r"\+(\d+)\+", label).group(1))
95
+ a3 = int(re.search(r"\+(\d+)/", label).group(1))
96
+ if re.search(r'\-([^\+]*)\+', labels[n + 1]).group(1) in ['sil', 'pau']:
97
+ a2_next = -1
98
+ else:
99
+ a2_next = int(
100
+ re.search(r"\+(\d+)\+", labels[n + 1]).group(1))
101
+ # Accent phrase boundary
102
+ if a3 == 1 and a2_next == 1:
103
+ text += ' '
104
+ # Falling
105
+ elif a1 == 0 and a2_next == a2 + 1:
106
+ text += '↓'
107
+ # Rising
108
+ elif a2 == 1 and a2_next == 2:
109
+ text += '↑'
110
+ if i < len(marks):
111
+ text += unidecode(marks[i]).replace(' ', '')
112
+ return text
113
+
114
+
115
+ def get_real_sokuon(text):
116
+ for regex, replacement in _real_sokuon:
117
+ text = re.sub(regex, replacement, text)
118
+ return text
119
+
120
+
121
+ def get_real_hatsuon(text):
122
+ for regex, replacement in _real_hatsuon:
123
+ text = re.sub(regex, replacement, text)
124
+ return text
125
+
126
+
127
+ def japanese_to_ipa(text):
128
+ text = japanese_to_romaji_with_accent(text).replace('...', '…')
129
+ text = re.sub(
130
+ r'([aiueo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
131
+ text = get_real_sokuon(text)
132
+ text = get_real_hatsuon(text)
133
+ for regex, replacement in _romaji_to_ipa:
134
+ text = re.sub(regex, replacement, text)
135
+ return text
136
+
137
+
138
+ def japanese_to_ipa2(text):
139
+ text = japanese_to_romaji_with_accent(text).replace('...', '…')
140
+ text = get_real_sokuon(text)
141
+ text = get_real_hatsuon(text)
142
+ for regex, replacement in _romaji_to_ipa2:
143
+ text = re.sub(regex, replacement, text)
144
+ return text
145
+
146
+
147
+ def japanese_to_ipa3(text):
148
+ text = japanese_to_ipa2(text).replace('n^', 'ȵ').replace(
149
+ 'ʃ', 'ɕ').replace('*', '\u0325').replace('#', '\u031a')
150
+ text = re.sub(
151
+ r'([aiɯeo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
152
+ text = re.sub(r'((?:^|\s)(?:ts|tɕ|[kpt]))', r'\1ʰ', text)
153
+ return text
text/mandarin.py ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import sys
3
+ import re
4
+ from pypinyin import lazy_pinyin, BOPOMOFO
5
+ import jieba
6
+ import cn2an
7
+
8
+
9
+ # List of (Latin alphabet, bopomofo) pairs:
10
+ _latin_to_bopomofo = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
11
+ ('a', 'ㄟˉ'),
12
+ ('b', 'ㄅㄧˋ'),
13
+ ('c', 'ㄙㄧˉ'),
14
+ ('d', 'ㄉㄧˋ'),
15
+ ('e', 'ㄧˋ'),
16
+ ('f', 'ㄝˊㄈㄨˋ'),
17
+ ('g', 'ㄐㄧˋ'),
18
+ ('h', 'ㄝˇㄑㄩˋ'),
19
+ ('i', 'ㄞˋ'),
20
+ ('j', 'ㄐㄟˋ'),
21
+ ('k', 'ㄎㄟˋ'),
22
+ ('l', 'ㄝˊㄛˋ'),
23
+ ('m', 'ㄝˊㄇㄨˋ'),
24
+ ('n', 'ㄣˉ'),
25
+ ('o', 'ㄡˉ'),
26
+ ('p', 'ㄆㄧˉ'),
27
+ ('q', 'ㄎㄧㄡˉ'),
28
+ ('r', 'ㄚˋ'),
29
+ ('s', 'ㄝˊㄙˋ'),
30
+ ('t', 'ㄊㄧˋ'),
31
+ ('u', 'ㄧㄡˉ'),
32
+ ('v', 'ㄨㄧˉ'),
33
+ ('w', 'ㄉㄚˋㄅㄨˋㄌㄧㄡˋ'),
34
+ ('x', 'ㄝˉㄎㄨˋㄙˋ'),
35
+ ('y', 'ㄨㄞˋ'),
36
+ ('z', 'ㄗㄟˋ')
37
+ ]]
38
+
39
+ # List of (bopomofo, romaji) pairs:
40
+ _bopomofo_to_romaji = [(re.compile('%s' % x[0]), x[1]) for x in [
41
+ ('ㄅㄛ', 'p⁼wo'),
42
+ ('ㄆㄛ', 'pʰwo'),
43
+ ('ㄇㄛ', 'mwo'),
44
+ ('ㄈㄛ', 'fwo'),
45
+ ('ㄅ', 'p⁼'),
46
+ ('ㄆ', 'pʰ'),
47
+ ('ㄇ', 'm'),
48
+ ('ㄈ', 'f'),
49
+ ('ㄉ', 't⁼'),
50
+ ('ㄊ', 'tʰ'),
51
+ ('ㄋ', 'n'),
52
+ ('ㄌ', 'l'),
53
+ ('ㄍ', 'k⁼'),
54
+ ('ㄎ', 'kʰ'),
55
+ ('ㄏ', 'h'),
56
+ ('ㄐ', 'ʧ⁼'),
57
+ ('ㄑ', 'ʧʰ'),
58
+ ('ㄒ', 'ʃ'),
59
+ ('ㄓ', 'ʦ`⁼'),
60
+ ('ㄔ', 'ʦ`ʰ'),
61
+ ('ㄕ', 's`'),
62
+ ('ㄖ', 'ɹ`'),
63
+ ('ㄗ', 'ʦ⁼'),
64
+ ('ㄘ', 'ʦʰ'),
65
+ ('ㄙ', 's'),
66
+ ('ㄚ', 'a'),
67
+ ('ㄛ', 'o'),
68
+ ('ㄜ', 'ə'),
69
+ ('ㄝ', 'e'),
70
+ ('ㄞ', 'ai'),
71
+ ('ㄟ', 'ei'),
72
+ ('ㄠ', 'au'),
73
+ ('ㄡ', 'ou'),
74
+ ('ㄧㄢ', 'yeNN'),
75
+ ('ㄢ', 'aNN'),
76
+ ('ㄧㄣ', 'iNN'),
77
+ ('ㄣ', 'əNN'),
78
+ ('ㄤ', 'aNg'),
79
+ ('ㄧㄥ', 'iNg'),
80
+ ('ㄨㄥ', 'uNg'),
81
+ ('ㄩㄥ', 'yuNg'),
82
+ ('ㄥ', 'əNg'),
83
+ ('ㄦ', 'əɻ'),
84
+ ('ㄧ', 'i'),
85
+ ('ㄨ', 'u'),
86
+ ('ㄩ', 'ɥ'),
87
+ ('ˉ', '→'),
88
+ ('ˊ', '↑'),
89
+ ('ˇ', '↓↑'),
90
+ ('ˋ', '↓'),
91
+ ('˙', ''),
92
+ (',', ','),
93
+ ('。', '.'),
94
+ ('!', '!'),
95
+ ('?', '?'),
96
+ ('—', '-')
97
+ ]]
98
+
99
+ # List of (romaji, ipa) pairs:
100
+ _romaji_to_ipa = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
101
+ ('ʃy', 'ʃ'),
102
+ ('ʧʰy', 'ʧʰ'),
103
+ ('ʧ⁼y', 'ʧ⁼'),
104
+ ('NN', 'n'),
105
+ ('Ng', 'ŋ'),
106
+ ('y', 'j'),
107
+ ('h', 'x')
108
+ ]]
109
+
110
+ # List of (bopomofo, ipa) pairs:
111
+ _bopomofo_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
112
+ ('ㄅㄛ', 'p⁼wo'),
113
+ ('ㄆㄛ', 'pʰwo'),
114
+ ('ㄇㄛ', 'mwo'),
115
+ ('ㄈㄛ', 'fwo'),
116
+ ('ㄅ', 'p⁼'),
117
+ ('ㄆ', 'pʰ'),
118
+ ('ㄇ', 'm'),
119
+ ('ㄈ', 'f'),
120
+ ('ㄉ', 't⁼'),
121
+ ('ㄊ', 'tʰ'),
122
+ ('ㄋ', 'n'),
123
+ ('ㄌ', 'l'),
124
+ ('ㄍ', 'k⁼'),
125
+ ('ㄎ', 'kʰ'),
126
+ ('ㄏ', 'x'),
127
+ ('ㄐ', 'tʃ⁼'),
128
+ ('ㄑ', 'tʃʰ'),
129
+ ('ㄒ', 'ʃ'),
130
+ ('ㄓ', 'ts`⁼'),
131
+ ('ㄔ', 'ts`ʰ'),
132
+ ('ㄕ', 's`'),
133
+ ('ㄖ', 'ɹ`'),
134
+ ('ㄗ', 'ts⁼'),
135
+ ('ㄘ', 'tsʰ'),
136
+ ('ㄙ', 's'),
137
+ ('ㄚ', 'a'),
138
+ ('ㄛ', 'o'),
139
+ ('ㄜ', 'ə'),
140
+ ('ㄝ', 'ɛ'),
141
+ ('ㄞ', 'aɪ'),
142
+ ('ㄟ', 'eɪ'),
143
+ ('ㄠ', 'ɑʊ'),
144
+ ('ㄡ', 'oʊ'),
145
+ ('ㄧㄢ', 'jɛn'),
146
+ ('ㄩㄢ', 'ɥæn'),
147
+ ('ㄢ', 'an'),
148
+ ('ㄧㄣ', 'in'),
149
+ ('ㄩㄣ', 'ɥn'),
150
+ ('ㄣ', 'ən'),
151
+ ('ㄤ', 'ɑŋ'),
152
+ ('ㄧㄥ', 'iŋ'),
153
+ ('ㄨㄥ', 'ʊŋ'),
154
+ ('ㄩㄥ', 'jʊŋ'),
155
+ ('ㄥ', 'əŋ'),
156
+ ('ㄦ', 'əɻ'),
157
+ ('ㄧ', 'i'),
158
+ ('ㄨ', 'u'),
159
+ ('ㄩ', 'ɥ'),
160
+ ('ˉ', '→'),
161
+ ('ˊ', '↑'),
162
+ ('ˇ', '↓↑'),
163
+ ('ˋ', '↓'),
164
+ ('˙', ''),
165
+ (',', ','),
166
+ ('。', '.'),
167
+ ('!', '!'),
168
+ ('?', '?'),
169
+ ('—', '-')
170
+ ]]
171
+
172
+ # List of (bopomofo, ipa2) pairs:
173
+ _bopomofo_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
174
+ ('ㄅㄛ', 'pwo'),
175
+ ('ㄆㄛ', 'pʰwo'),
176
+ ('ㄇㄛ', 'mwo'),
177
+ ('ㄈㄛ', 'fwo'),
178
+ ('ㄅ', 'p'),
179
+ ('ㄆ', 'pʰ'),
180
+ ('ㄇ', 'm'),
181
+ ('ㄈ', 'f'),
182
+ ('ㄉ', 't'),
183
+ ('ㄊ', 'tʰ'),
184
+ ('ㄋ', 'n'),
185
+ ('ㄌ', 'l'),
186
+ ('ㄍ', 'k'),
187
+ ('ㄎ', 'kʰ'),
188
+ ('ㄏ', 'h'),
189
+ ('ㄐ', 'tɕ'),
190
+ ('ㄑ', 'tɕʰ'),
191
+ ('ㄒ', 'ɕ'),
192
+ ('ㄓ', 'tʂ'),
193
+ ('ㄔ', 'tʂʰ'),
194
+ ('ㄕ', 'ʂ'),
195
+ ('ㄖ', 'ɻ'),
196
+ ('ㄗ', 'ts'),
197
+ ('ㄘ', 'tsʰ'),
198
+ ('ㄙ', 's'),
199
+ ('ㄚ', 'a'),
200
+ ('ㄛ', 'o'),
201
+ ('ㄜ', 'ɤ'),
202
+ ('ㄝ', 'ɛ'),
203
+ ('ㄞ', 'aɪ'),
204
+ ('ㄟ', 'eɪ'),
205
+ ('ㄠ', 'ɑʊ'),
206
+ ('ㄡ', 'oʊ'),
207
+ ('ㄧㄢ', 'jɛn'),
208
+ ('ㄩㄢ', 'yæn'),
209
+ ('ㄢ', 'an'),
210
+ ('ㄧㄣ', 'in'),
211
+ ('ㄩㄣ', 'yn'),
212
+ ('ㄣ', 'ən'),
213
+ ('ㄤ', 'ɑŋ'),
214
+ ('ㄧㄥ', 'iŋ'),
215
+ ('ㄨㄥ', 'ʊŋ'),
216
+ ('ㄩㄥ', 'jʊŋ'),
217
+ ('ㄥ', 'ɤŋ'),
218
+ ('ㄦ', 'əɻ'),
219
+ ('ㄧ', 'i'),
220
+ ('ㄨ', 'u'),
221
+ ('ㄩ', 'y'),
222
+ ('ˉ', '˥'),
223
+ ('ˊ', '˧˥'),
224
+ ('ˇ', '˨˩˦'),
225
+ ('ˋ', '˥˩'),
226
+ ('˙', ''),
227
+ (',', ','),
228
+ ('。', '.'),
229
+ ('!', '!'),
230
+ ('?', '?'),
231
+ ('—', '-')
232
+ ]]
233
+
234
+
235
+ def number_to_chinese(text):
236
+ numbers = re.findall(r'\d+(?:\.?\d+)?', text)
237
+ for number in numbers:
238
+ text = text.replace(number, cn2an.an2cn(number), 1)
239
+ return text
240
+
241
+
242
+ def chinese_to_bopomofo(text, taiwanese=False):
243
+ text = text.replace('、', ',').replace(';', ',').replace(':', ',')
244
+ words = jieba.lcut(text, cut_all=False)
245
+ text = ''
246
+ for word in words:
247
+ bopomofos = lazy_pinyin(word, BOPOMOFO)
248
+ if not re.search('[\u4e00-\u9fff]', word):
249
+ text += word
250
+ continue
251
+ for i in range(len(bopomofos)):
252
+ bopomofos[i] = re.sub(r'([\u3105-\u3129])$', r'\1ˉ', bopomofos[i])
253
+ if text != '':
254
+ text += ' '
255
+ if taiwanese:
256
+ text += '#'+'#'.join(bopomofos)
257
+ else:
258
+ text += ''.join(bopomofos)
259
+ return text
260
+
261
+
262
+ def latin_to_bopomofo(text):
263
+ for regex, replacement in _latin_to_bopomofo:
264
+ text = re.sub(regex, replacement, text)
265
+ return text
266
+
267
+
268
+ def bopomofo_to_romaji(text):
269
+ for regex, replacement in _bopomofo_to_romaji:
270
+ text = re.sub(regex, replacement, text)
271
+ return text
272
+
273
+
274
+ def bopomofo_to_ipa(text):
275
+ for regex, replacement in _bopomofo_to_ipa:
276
+ text = re.sub(regex, replacement, text)
277
+ return text
278
+
279
+
280
+ def bopomofo_to_ipa2(text):
281
+ for regex, replacement in _bopomofo_to_ipa2:
282
+ text = re.sub(regex, replacement, text)
283
+ return text
284
+
285
+
286
+ def chinese_to_romaji(text):
287
+ text = number_to_chinese(text)
288
+ text = chinese_to_bopomofo(text)
289
+ text = latin_to_bopomofo(text)
290
+ text = bopomofo_to_romaji(text)
291
+ text = re.sub('i([aoe])', r'y\1', text)
292
+ text = re.sub('u([aoəe])', r'w\1', text)
293
+ text = re.sub('([ʦsɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
294
+ r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
295
+ text = re.sub('([ʦs][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
296
+ return text
297
+
298
+
299
+ def chinese_to_lazy_ipa(text):
300
+ text = chinese_to_romaji(text)
301
+ for regex, replacement in _romaji_to_ipa:
302
+ text = re.sub(regex, replacement, text)
303
+ return text
304
+
305
+
306
+ def chinese_to_ipa(text):
307
+ text = number_to_chinese(text)
308
+ text = chinese_to_bopomofo(text)
309
+ text = latin_to_bopomofo(text)
310
+ text = bopomofo_to_ipa(text)
311
+ text = re.sub('i([aoe])', r'j\1', text)
312
+ text = re.sub('u([aoəe])', r'w\1', text)
313
+ text = re.sub('([sɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
314
+ r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
315
+ text = re.sub('([s][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
316
+ return text
317
+
318
+
319
+ def chinese_to_ipa2(text, taiwanese=False):
320
+ text = number_to_chinese(text)
321
+ text = chinese_to_bopomofo(text, taiwanese)
322
+ text = latin_to_bopomofo(text)
323
+ text = bopomofo_to_ipa2(text)
324
+ text = re.sub(r'i([aoe])', r'j\1', text)
325
+ text = re.sub(r'u([aoəe])', r'w\1', text)
326
+ text = re.sub(r'([ʂɹ]ʰ?)([˩˨˧˦˥ ]+|$)', r'\1ʅ\2', text)
327
+ text = re.sub(r'(sʰ?)([˩˨˧˦˥ ]+|$)', r'\1ɿ\2', text)
328
+ return text
text/symbols.py ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ '''
2
+ Defines the set of symbols used in text input to the model.
3
+ '''
4
+ _pad = '_'
5
+ _punctuation = ';:,.!?¡¿—…"«»“” '
6
+ _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
7
+ _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
8
+
9
+
10
+ # Export all symbols:
11
+ symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
12
+ # Special symbol ids
13
+ SPACE_ID = symbols.index(" ")
transforms.py ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+
4
+ import numpy as np
5
+
6
+
7
+ DEFAULT_MIN_BIN_WIDTH = 1e-3
8
+ DEFAULT_MIN_BIN_HEIGHT = 1e-3
9
+ DEFAULT_MIN_DERIVATIVE = 1e-3
10
+
11
+
12
+ def piecewise_rational_quadratic_transform(inputs,
13
+ unnormalized_widths,
14
+ unnormalized_heights,
15
+ unnormalized_derivatives,
16
+ inverse=False,
17
+ tails=None,
18
+ tail_bound=1.,
19
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
20
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
21
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
22
+
23
+ if tails is None:
24
+ spline_fn = rational_quadratic_spline
25
+ spline_kwargs = {}
26
+ else:
27
+ spline_fn = unconstrained_rational_quadratic_spline
28
+ spline_kwargs = {
29
+ 'tails': tails,
30
+ 'tail_bound': tail_bound
31
+ }
32
+
33
+ outputs, logabsdet = spline_fn(
34
+ inputs=inputs,
35
+ unnormalized_widths=unnormalized_widths,
36
+ unnormalized_heights=unnormalized_heights,
37
+ unnormalized_derivatives=unnormalized_derivatives,
38
+ inverse=inverse,
39
+ min_bin_width=min_bin_width,
40
+ min_bin_height=min_bin_height,
41
+ min_derivative=min_derivative,
42
+ **spline_kwargs
43
+ )
44
+ return outputs, logabsdet
45
+
46
+
47
+ def searchsorted(bin_locations, inputs, eps=1e-6):
48
+ bin_locations[..., -1] += eps
49
+ return torch.sum(
50
+ inputs[..., None] >= bin_locations,
51
+ dim=-1
52
+ ) - 1
53
+
54
+
55
+ def unconstrained_rational_quadratic_spline(inputs,
56
+ unnormalized_widths,
57
+ unnormalized_heights,
58
+ unnormalized_derivatives,
59
+ inverse=False,
60
+ tails='linear',
61
+ tail_bound=1.,
62
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
63
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
64
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
65
+ inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
66
+ outside_interval_mask = ~inside_interval_mask
67
+
68
+ outputs = torch.zeros_like(inputs)
69
+ logabsdet = torch.zeros_like(inputs)
70
+
71
+ if tails == 'linear':
72
+ unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
73
+ constant = np.log(np.exp(1 - min_derivative) - 1)
74
+ unnormalized_derivatives[..., 0] = constant
75
+ unnormalized_derivatives[..., -1] = constant
76
+
77
+ outputs[outside_interval_mask] = inputs[outside_interval_mask]
78
+ logabsdet[outside_interval_mask] = 0
79
+ else:
80
+ raise RuntimeError('{} tails are not implemented.'.format(tails))
81
+
82
+ outputs[inside_interval_mask], logabsdet[inside_interval_mask] = rational_quadratic_spline(
83
+ inputs=inputs[inside_interval_mask],
84
+ unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
85
+ unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
86
+ unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
87
+ inverse=inverse,
88
+ left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound,
89
+ min_bin_width=min_bin_width,
90
+ min_bin_height=min_bin_height,
91
+ min_derivative=min_derivative
92
+ )
93
+
94
+ return outputs, logabsdet
95
+
96
+ def rational_quadratic_spline(inputs,
97
+ unnormalized_widths,
98
+ unnormalized_heights,
99
+ unnormalized_derivatives,
100
+ inverse=False,
101
+ left=0., right=1., bottom=0., top=1.,
102
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
103
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
104
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
105
+ if torch.min(inputs) < left or torch.max(inputs) > right:
106
+ raise ValueError('Input to a transform is not within its domain')
107
+
108
+ num_bins = unnormalized_widths.shape[-1]
109
+
110
+ if min_bin_width * num_bins > 1.0:
111
+ raise ValueError('Minimal bin width too large for the number of bins')
112
+ if min_bin_height * num_bins > 1.0:
113
+ raise ValueError('Minimal bin height too large for the number of bins')
114
+
115
+ widths = F.softmax(unnormalized_widths, dim=-1)
116
+ widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
117
+ cumwidths = torch.cumsum(widths, dim=-1)
118
+ cumwidths = F.pad(cumwidths, pad=(1, 0), mode='constant', value=0.0)
119
+ cumwidths = (right - left) * cumwidths + left
120
+ cumwidths[..., 0] = left
121
+ cumwidths[..., -1] = right
122
+ widths = cumwidths[..., 1:] - cumwidths[..., :-1]
123
+
124
+ derivatives = min_derivative + F.softplus(unnormalized_derivatives)
125
+
126
+ heights = F.softmax(unnormalized_heights, dim=-1)
127
+ heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
128
+ cumheights = torch.cumsum(heights, dim=-1)
129
+ cumheights = F.pad(cumheights, pad=(1, 0), mode='constant', value=0.0)
130
+ cumheights = (top - bottom) * cumheights + bottom
131
+ cumheights[..., 0] = bottom
132
+ cumheights[..., -1] = top
133
+ heights = cumheights[..., 1:] - cumheights[..., :-1]
134
+
135
+ if inverse:
136
+ bin_idx = searchsorted(cumheights, inputs)[..., None]
137
+ else:
138
+ bin_idx = searchsorted(cumwidths, inputs)[..., None]
139
+
140
+ input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
141
+ input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
142
+
143
+ input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
144
+ delta = heights / widths
145
+ input_delta = delta.gather(-1, bin_idx)[..., 0]
146
+
147
+ input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
148
+ input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
149
+
150
+ input_heights = heights.gather(-1, bin_idx)[..., 0]
151
+
152
+ if inverse:
153
+ a = (((inputs - input_cumheights) * (input_derivatives
154
+ + input_derivatives_plus_one
155
+ - 2 * input_delta)
156
+ + input_heights * (input_delta - input_derivatives)))
157
+ b = (input_heights * input_derivatives
158
+ - (inputs - input_cumheights) * (input_derivatives
159
+ + input_derivatives_plus_one
160
+ - 2 * input_delta))
161
+ c = - input_delta * (inputs - input_cumheights)
162
+
163
+ discriminant = b.pow(2) - 4 * a * c
164
+ assert (discriminant >= 0).all()
165
+
166
+ root = (2 * c) / (-b - torch.sqrt(discriminant))
167
+ outputs = root * input_bin_widths + input_cumwidths
168
+
169
+ theta_one_minus_theta = root * (1 - root)
170
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
171
+ * theta_one_minus_theta)
172
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * root.pow(2)
173
+ + 2 * input_delta * theta_one_minus_theta
174
+ + input_derivatives * (1 - root).pow(2))
175
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
176
+
177
+ return outputs, -logabsdet
178
+ else:
179
+ theta = (inputs - input_cumwidths) / input_bin_widths
180
+ theta_one_minus_theta = theta * (1 - theta)
181
+
182
+ numerator = input_heights * (input_delta * theta.pow(2)
183
+ + input_derivatives * theta_one_minus_theta)
184
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
185
+ * theta_one_minus_theta)
186
+ outputs = input_cumheights + numerator / denominator
187
+
188
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2)
189
+ + 2 * input_delta * theta_one_minus_theta
190
+ + input_derivatives * (1 - theta).pow(2))
191
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
192
+
193
+ return outputs, logabsdet
utils.py ADDED
@@ -0,0 +1,263 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import glob
3
+ import sys
4
+ import argparse
5
+ import logging
6
+ import json
7
+ import subprocess
8
+ import numpy as np
9
+ from scipy.io.wavfile import read
10
+ import torch
11
+
12
+ MATPLOTLIB_FLAG = False
13
+
14
+ logging.basicConfig(stream=sys.stdout, level=logging.WARNING)
15
+ logger = logging
16
+
17
+
18
+ def load_checkpoint(checkpoint_path, model, optimizer=None):
19
+ assert os.path.isfile(checkpoint_path)
20
+ checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
21
+ iteration = checkpoint_dict['iteration']
22
+ learning_rate = checkpoint_dict['learning_rate']
23
+ if optimizer is not None:
24
+ optimizer.load_state_dict(checkpoint_dict['optimizer'])
25
+ saved_state_dict = checkpoint_dict['model']
26
+ if hasattr(model, 'module'):
27
+ state_dict = model.module.state_dict()
28
+ else:
29
+ state_dict = model.state_dict()
30
+ new_state_dict= {}
31
+ for k, v in state_dict.items():
32
+ try:
33
+ new_state_dict[k] = saved_state_dict[k]
34
+ except:
35
+ logger.info("%s is not in the checkpoint" % k)
36
+ new_state_dict[k] = v
37
+ if hasattr(model, 'module'):
38
+ model.module.load_state_dict(new_state_dict)
39
+ else:
40
+ model.load_state_dict(new_state_dict)
41
+ logger.info("Loaded checkpoint '{}' (iteration {})" .format(
42
+ checkpoint_path, iteration))
43
+ return model, optimizer, learning_rate, iteration
44
+
45
+
46
+ def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
47
+ ckptname = checkpoint_path.split("/")[-1]
48
+ newest_step = int(ckptname.split(".")[0].split("_")[1])
49
+ last_ckptname = checkpoint_path.replace(str(newest_step), str(newest_step-3000))
50
+ if newest_step >= 3000:
51
+ os.system(f"rm {last_ckptname}")
52
+ logger.info("Saving model and optimizer state at iteration {} to {}".format(
53
+ iteration, checkpoint_path))
54
+ if hasattr(model, 'module'):
55
+ state_dict = model.module.state_dict()
56
+ else:
57
+ state_dict = model.state_dict()
58
+ torch.save({'model': state_dict,
59
+ 'iteration': iteration,
60
+ 'optimizer': optimizer.state_dict(),
61
+ 'learning_rate': learning_rate}, checkpoint_path)
62
+
63
+
64
+ def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050):
65
+ for k, v in scalars.items():
66
+ writer.add_scalar(k, v, global_step)
67
+ for k, v in histograms.items():
68
+ writer.add_histogram(k, v, global_step)
69
+ for k, v in images.items():
70
+ writer.add_image(k, v, global_step, dataformats='HWC')
71
+ for k, v in audios.items():
72
+ writer.add_audio(k, v, global_step, audio_sampling_rate)
73
+
74
+
75
+ def latest_checkpoint_path(dir_path, regex="G_*.pth"):
76
+ f_list = glob.glob(os.path.join(dir_path, regex))
77
+ f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
78
+ x = f_list[-1]
79
+ print(x)
80
+ return x
81
+
82
+
83
+ def plot_spectrogram_to_numpy(spectrogram):
84
+ global MATPLOTLIB_FLAG
85
+ if not MATPLOTLIB_FLAG:
86
+ import matplotlib
87
+ matplotlib.use("Agg")
88
+ MATPLOTLIB_FLAG = True
89
+ mpl_logger = logging.getLogger('matplotlib')
90
+ mpl_logger.setLevel(logging.WARNING)
91
+ import matplotlib.pylab as plt
92
+ import numpy as np
93
+
94
+ fig, ax = plt.subplots(figsize=(10,2))
95
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
96
+ interpolation='none')
97
+ plt.colorbar(im, ax=ax)
98
+ plt.xlabel("Frames")
99
+ plt.ylabel("Channels")
100
+ plt.tight_layout()
101
+
102
+ fig.canvas.draw()
103
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
104
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
105
+ plt.close()
106
+ return data
107
+
108
+
109
+ def plot_alignment_to_numpy(alignment, info=None):
110
+ global MATPLOTLIB_FLAG
111
+ if not MATPLOTLIB_FLAG:
112
+ import matplotlib
113
+ matplotlib.use("Agg")
114
+ MATPLOTLIB_FLAG = True
115
+ mpl_logger = logging.getLogger('matplotlib')
116
+ mpl_logger.setLevel(logging.WARNING)
117
+ import matplotlib.pylab as plt
118
+ import numpy as np
119
+
120
+ fig, ax = plt.subplots(figsize=(6, 4))
121
+ im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
122
+ interpolation='none')
123
+ fig.colorbar(im, ax=ax)
124
+ xlabel = 'Decoder timestep'
125
+ if info is not None:
126
+ xlabel += '\n\n' + info
127
+ plt.xlabel(xlabel)
128
+ plt.ylabel('Encoder timestep')
129
+ plt.tight_layout()
130
+
131
+ fig.canvas.draw()
132
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
133
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
134
+ plt.close()
135
+ return data
136
+
137
+
138
+ def load_wav_to_torch(full_path):
139
+ sampling_rate, data = read(full_path)
140
+ return torch.FloatTensor(data.astype(np.float32)), sampling_rate
141
+
142
+
143
+ def load_filepaths_and_text(filename, split="|"):
144
+ with open(filename, encoding='utf-8') as f:
145
+ filepaths_and_text = [line.strip().split(split) for line in f]
146
+ return filepaths_and_text
147
+
148
+
149
+ def get_hparams(init=True):
150
+ parser = argparse.ArgumentParser()
151
+ parser.add_argument('-c', '--config', type=str, default="./configs/base.json",
152
+ help='JSON file for configuration')
153
+ parser.add_argument('-m', '--model', type=str, required=True,
154
+ help='Model name')
155
+
156
+ args = parser.parse_args()
157
+ model_dir = os.path.join("./logs", args.model)
158
+
159
+ if not os.path.exists(model_dir):
160
+ os.makedirs(model_dir)
161
+
162
+ config_path = args.config
163
+ config_save_path = os.path.join(model_dir, "config.json")
164
+ if init:
165
+ with open(config_path, "r") as f:
166
+ data = f.read()
167
+ with open(config_save_path, "w") as f:
168
+ f.write(data)
169
+ else:
170
+ with open(config_save_path, "r") as f:
171
+ data = f.read()
172
+ config = json.loads(data)
173
+
174
+ hparams = HParams(**config)
175
+ hparams.model_dir = model_dir
176
+ return hparams
177
+
178
+
179
+ def get_hparams_from_dir(model_dir):
180
+ config_save_path = os.path.join(model_dir, "config.json")
181
+ with open(config_save_path, "r") as f:
182
+ data = f.read()
183
+ config = json.loads(data)
184
+
185
+ hparams =HParams(**config)
186
+ hparams.model_dir = model_dir
187
+ return hparams
188
+
189
+
190
+ def get_hparams_from_file(config_path):
191
+ with open(config_path, "r") as f:
192
+ data = f.read()
193
+ config = json.loads(data)
194
+
195
+ hparams =HParams(**config)
196
+ return hparams
197
+
198
+
199
+ def check_git_hash(model_dir):
200
+ source_dir = os.path.dirname(os.path.realpath(__file__))
201
+ if not os.path.exists(os.path.join(source_dir, ".git")):
202
+ logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
203
+ source_dir
204
+ ))
205
+ return
206
+
207
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
208
+
209
+ path = os.path.join(model_dir, "githash")
210
+ if os.path.exists(path):
211
+ saved_hash = open(path).read()
212
+ if saved_hash != cur_hash:
213
+ logger.warn("git hash values are different. {}(saved) != {}(current)".format(
214
+ saved_hash[:8], cur_hash[:8]))
215
+ else:
216
+ open(path, "w").write(cur_hash)
217
+
218
+
219
+ def get_logger(model_dir, filename="train.log"):
220
+ global logger
221
+ logger = logging.getLogger(os.path.basename(model_dir))
222
+ logger.setLevel(logging.DEBUG)
223
+
224
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
225
+ if not os.path.exists(model_dir):
226
+ os.makedirs(model_dir)
227
+ h = logging.FileHandler(os.path.join(model_dir, filename))
228
+ h.setLevel(logging.DEBUG)
229
+ h.setFormatter(formatter)
230
+ logger.addHandler(h)
231
+ return logger
232
+
233
+
234
+ class HParams():
235
+ def __init__(self, **kwargs):
236
+ for k, v in kwargs.items():
237
+ if type(v) == dict:
238
+ v = HParams(**v)
239
+ self[k] = v
240
+
241
+ def keys(self):
242
+ return self.__dict__.keys()
243
+
244
+ def items(self):
245
+ return self.__dict__.items()
246
+
247
+ def values(self):
248
+ return self.__dict__.values()
249
+
250
+ def __len__(self):
251
+ return len(self.__dict__)
252
+
253
+ def __getitem__(self, key):
254
+ return getattr(self, key)
255
+
256
+ def __setitem__(self, key, value):
257
+ return setattr(self, key, value)
258
+
259
+ def __contains__(self, key):
260
+ return key in self.__dict__
261
+
262
+ def __repr__(self):
263
+ return self.__dict__.__repr__()