zhzluke96
commited on
Commit
•
84cfd61
1
Parent(s):
22884c9
update
Browse files- modules/devices.py +8 -0
- modules/generate_audio.py +4 -0
- modules/normalization.py +38 -10
- modules/utils/audio.py +10 -0
- webui.py +27 -13
modules/devices.py
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
|
4 |
+
def torch_gc():
|
5 |
+
if torch.cuda.is_available():
|
6 |
+
with torch.cuda.device("cuda"):
|
7 |
+
torch.cuda.empty_cache()
|
8 |
+
torch.cuda.ipc_collect()
|
modules/generate_audio.py
CHANGED
@@ -8,6 +8,8 @@ from modules import models, config
|
|
8 |
|
9 |
import logging
|
10 |
|
|
|
|
|
11 |
logger = logging.getLogger(__name__)
|
12 |
|
13 |
|
@@ -96,6 +98,8 @@ def generate_audio_batch(
|
|
96 |
|
97 |
sample_rate = 24000
|
98 |
|
|
|
|
|
99 |
return [(sample_rate, np.array(wav).flatten().astype(np.float32)) for wav in wavs]
|
100 |
|
101 |
|
|
|
8 |
|
9 |
import logging
|
10 |
|
11 |
+
from modules import devices
|
12 |
+
|
13 |
logger = logging.getLogger(__name__)
|
14 |
|
15 |
|
|
|
98 |
|
99 |
sample_rate = 24000
|
100 |
|
101 |
+
devices.torch_gc()
|
102 |
+
|
103 |
return [(sample_rate, np.array(wav).flatten().astype(np.float32)) for wav in wavs]
|
104 |
|
105 |
|
modules/normalization.py
CHANGED
@@ -75,13 +75,15 @@ character_map = {
|
|
75 |
"“": " ",
|
76 |
"’": " ",
|
77 |
"”": " ",
|
|
|
|
|
78 |
":": ",",
|
79 |
";": ",",
|
80 |
"!": ".",
|
81 |
"(": ",",
|
82 |
")": ",",
|
83 |
-
|
84 |
-
|
85 |
">": ",",
|
86 |
"<": ",",
|
87 |
"-": ",",
|
@@ -110,13 +112,6 @@ def apply_emoji_map(text):
|
|
110 |
return emojiswitch.demojize(text, delimiters=("", ""), lang="zh")
|
111 |
|
112 |
|
113 |
-
@pre_normalize()
|
114 |
-
def apply_markdown_to_text(text):
|
115 |
-
if is_markdown(text):
|
116 |
-
text = markdown_to_text(text)
|
117 |
-
return text
|
118 |
-
|
119 |
-
|
120 |
@post_normalize()
|
121 |
def insert_spaces_between_uppercase(s):
|
122 |
# 使用正则表达式在每个相邻的大写字母之间插入空格
|
@@ -127,6 +122,29 @@ def insert_spaces_between_uppercase(s):
|
|
127 |
)
|
128 |
|
129 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
def ensure_suffix(a: str, b: str, c: str):
|
131 |
a = a.strip()
|
132 |
if not a.endswith(b):
|
@@ -171,6 +189,7 @@ def sentence_normalize(sentence_text: str):
|
|
171 |
sentences = tx.normalize(part)
|
172 |
dest_text = ""
|
173 |
for sentence in sentences:
|
|
|
174 |
dest_text += sentence
|
175 |
return dest_text
|
176 |
|
@@ -197,7 +216,6 @@ def text_normalize(text, is_end=False):
|
|
197 |
lines = [line for line in lines if line]
|
198 |
lines = [sentence_normalize(line) for line in lines]
|
199 |
content = "\n".join(lines)
|
200 |
-
content = apply_post_normalize(content)
|
201 |
return content
|
202 |
|
203 |
|
@@ -216,6 +234,16 @@ console.log('1')
|
|
216 |
|
217 |
*一条文本*
|
218 |
""",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
]
|
220 |
|
221 |
for i, test_case in enumerate(test_cases):
|
|
|
75 |
"“": " ",
|
76 |
"’": " ",
|
77 |
"”": " ",
|
78 |
+
'"': " ",
|
79 |
+
"'": " ",
|
80 |
":": ",",
|
81 |
";": ",",
|
82 |
"!": ".",
|
83 |
"(": ",",
|
84 |
")": ",",
|
85 |
+
"[": ",",
|
86 |
+
"]": ",",
|
87 |
">": ",",
|
88 |
"<": ",",
|
89 |
"-": ",",
|
|
|
112 |
return emojiswitch.demojize(text, delimiters=("", ""), lang="zh")
|
113 |
|
114 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
@post_normalize()
|
116 |
def insert_spaces_between_uppercase(s):
|
117 |
# 使用正则表达式在每个相邻的大写字母之间插入空格
|
|
|
122 |
)
|
123 |
|
124 |
|
125 |
+
@pre_normalize()
|
126 |
+
def apply_markdown_to_text(text):
|
127 |
+
if is_markdown(text):
|
128 |
+
text = markdown_to_text(text)
|
129 |
+
return text
|
130 |
+
|
131 |
+
|
132 |
+
# 将 "xxx" => \nxxx\n
|
133 |
+
# 将 'xxx' => \nxxx\n
|
134 |
+
@pre_normalize()
|
135 |
+
def replace_quotes(text):
|
136 |
+
repl = r"\n\1\n"
|
137 |
+
patterns = [
|
138 |
+
['"', '"'],
|
139 |
+
["'", "'"],
|
140 |
+
["“", "”"],
|
141 |
+
["‘", "’"],
|
142 |
+
]
|
143 |
+
for p in patterns:
|
144 |
+
text = re.sub(rf"({p[0]}[^{p[0]}{p[1]}]+?{p[1]})", repl, text)
|
145 |
+
return text
|
146 |
+
|
147 |
+
|
148 |
def ensure_suffix(a: str, b: str, c: str):
|
149 |
a = a.strip()
|
150 |
if not a.endswith(b):
|
|
|
189 |
sentences = tx.normalize(part)
|
190 |
dest_text = ""
|
191 |
for sentence in sentences:
|
192 |
+
sentence = apply_post_normalize(sentence)
|
193 |
dest_text += sentence
|
194 |
return dest_text
|
195 |
|
|
|
216 |
lines = [line for line in lines if line]
|
217 |
lines = [sentence_normalize(line) for line in lines]
|
218 |
content = "\n".join(lines)
|
|
|
219 |
return content
|
220 |
|
221 |
|
|
|
234 |
|
235 |
*一条文本*
|
236 |
""",
|
237 |
+
"""
|
238 |
+
在沙漠、岩石、雪地上行走了很长的时间以后,小王子终于发现了一条大路。所有的大路都是通往人住的地方的。
|
239 |
+
“你们好。”小王子说。
|
240 |
+
这是一个玫瑰盛开的花园。
|
241 |
+
“你好。”玫瑰花说道。
|
242 |
+
小王子瞅着这些花,它们全都和他的那朵花一样。
|
243 |
+
“你们是什么花?”小王子惊奇地问。
|
244 |
+
“我们是玫瑰花。”花儿们说道。
|
245 |
+
“啊!”小王子说……。
|
246 |
+
""",
|
247 |
]
|
248 |
|
249 |
for i, test_case in enumerate(test_cases):
|
modules/utils/audio.py
CHANGED
@@ -5,6 +5,16 @@ import pyrubberband as pyrb
|
|
5 |
import numpy as np
|
6 |
from io import BytesIO
|
7 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
def audiosegment_to_librosawav(audiosegment):
|
10 |
channel_sounds = audiosegment.split_to_mono()
|
|
|
5 |
import numpy as np
|
6 |
from io import BytesIO
|
7 |
|
8 |
+
INT16_MAX = np.iinfo(np.int16).max
|
9 |
+
|
10 |
+
|
11 |
+
def audio_to_int16(audio_data):
|
12 |
+
if audio_data.dtype == np.float32:
|
13 |
+
audio_data = (audio_data * INT16_MAX).astype(np.int16)
|
14 |
+
if audio_data.dtype == np.float16:
|
15 |
+
audio_data = (audio_data * INT16_MAX).astype(np.int16)
|
16 |
+
return audio_data
|
17 |
+
|
18 |
|
19 |
def audiosegment_to_librosawav(audiosegment):
|
20 |
channel_sounds = audiosegment.split_to_mono()
|
webui.py
CHANGED
@@ -1,4 +1,16 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import os
|
3 |
import logging
|
4 |
|
@@ -29,7 +41,7 @@ from modules.api.utils import calc_spk_style
|
|
29 |
from modules.normalization import text_normalize
|
30 |
from modules import refiner, config
|
31 |
|
32 |
-
from modules.utils import env
|
33 |
from modules.SentenceSplitter import SentenceSplitter
|
34 |
|
35 |
torch._dynamo.config.cache_size_limit = 64
|
@@ -40,7 +52,7 @@ webui_config = {
|
|
40 |
"tts_max": 1000,
|
41 |
"ssml_max": 5000,
|
42 |
"spliter_threshold": 100,
|
43 |
-
"max_batch_size":
|
44 |
}
|
45 |
|
46 |
|
@@ -65,7 +77,7 @@ def segments_length_limit(segments, total_max: int):
|
|
65 |
|
66 |
@torch.inference_mode()
|
67 |
@spaces.GPU
|
68 |
-
def synthesize_ssml(ssml: str, batch_size=
|
69 |
try:
|
70 |
batch_size = int(batch_size)
|
71 |
except Exception:
|
@@ -92,7 +104,10 @@ def synthesize_ssml(ssml: str, batch_size=8):
|
|
92 |
|
93 |
buffer.seek(0)
|
94 |
|
95 |
-
|
|
|
|
|
|
|
96 |
|
97 |
|
98 |
@torch.inference_mode()
|
@@ -110,12 +125,12 @@ def tts_generate(
|
|
110 |
prefix,
|
111 |
style,
|
112 |
disable_normalize=False,
|
113 |
-
batch_size=
|
114 |
):
|
115 |
try:
|
116 |
batch_size = int(batch_size)
|
117 |
except Exception:
|
118 |
-
batch_size =
|
119 |
|
120 |
max_len = webui_config["tts_max"]
|
121 |
text = text.strip()[0:max_len]
|
@@ -157,8 +172,6 @@ def tts_generate(
|
|
157 |
prompt2=prompt2,
|
158 |
prefix=prefix,
|
159 |
)
|
160 |
-
|
161 |
-
return sample_rate, audio_data
|
162 |
else:
|
163 |
spliter = SentenceSplitter(webui_config["spliter_threshold"])
|
164 |
sentences = spliter.parse(text)
|
@@ -178,7 +191,8 @@ def tts_generate(
|
|
178 |
sample_rate = audio_data_batch[0][0]
|
179 |
audio_data = np.concatenate([data for _, data in audio_data_batch])
|
180 |
|
181 |
-
|
|
|
182 |
|
183 |
|
184 |
@torch.inference_mode()
|
@@ -366,7 +380,7 @@ def create_tts_interface():
|
|
366 |
batch_size_input = gr.Slider(
|
367 |
1,
|
368 |
webui_config["max_batch_size"],
|
369 |
-
value=
|
370 |
step=1,
|
371 |
label="Batch Size",
|
372 |
)
|
@@ -593,7 +607,7 @@ def create_ssml_interface():
|
|
593 |
# batch size
|
594 |
batch_size_input = gr.Slider(
|
595 |
label="Batch Size",
|
596 |
-
value=
|
597 |
minimum=1,
|
598 |
maximum=webui_config["max_batch_size"],
|
599 |
step=1,
|
@@ -892,7 +906,7 @@ if __name__ == "__main__":
|
|
892 |
|
893 |
webui_config["tts_max"] = env.get_env_or_arg(args, "tts_max_len", 1000, int)
|
894 |
webui_config["ssml_max"] = env.get_env_or_arg(args, "ssml_max_len", 5000, int)
|
895 |
-
webui_config["max_batch_size"] = env.get_env_or_arg(args, "max_batch_size",
|
896 |
|
897 |
demo = create_interface()
|
898 |
|
|
|
1 |
+
try:
|
2 |
+
import spaces
|
3 |
+
except:
|
4 |
+
|
5 |
+
class NoneSpaces:
|
6 |
+
def __init__(self):
|
7 |
+
pass
|
8 |
+
|
9 |
+
def GPU(self, fn):
|
10 |
+
return fn
|
11 |
+
|
12 |
+
spaces = NoneSpaces()
|
13 |
+
|
14 |
import os
|
15 |
import logging
|
16 |
|
|
|
41 |
from modules.normalization import text_normalize
|
42 |
from modules import refiner, config
|
43 |
|
44 |
+
from modules.utils import env, audio
|
45 |
from modules.SentenceSplitter import SentenceSplitter
|
46 |
|
47 |
torch._dynamo.config.cache_size_limit = 64
|
|
|
52 |
"tts_max": 1000,
|
53 |
"ssml_max": 5000,
|
54 |
"spliter_threshold": 100,
|
55 |
+
"max_batch_size": 8,
|
56 |
}
|
57 |
|
58 |
|
|
|
77 |
|
78 |
@torch.inference_mode()
|
79 |
@spaces.GPU
|
80 |
+
def synthesize_ssml(ssml: str, batch_size=4):
|
81 |
try:
|
82 |
batch_size = int(batch_size)
|
83 |
except Exception:
|
|
|
104 |
|
105 |
buffer.seek(0)
|
106 |
|
107 |
+
audio_data = buffer.read()
|
108 |
+
audio_data = audio.audio_to_int16(audio_data)
|
109 |
+
|
110 |
+
return audio_data
|
111 |
|
112 |
|
113 |
@torch.inference_mode()
|
|
|
125 |
prefix,
|
126 |
style,
|
127 |
disable_normalize=False,
|
128 |
+
batch_size=4,
|
129 |
):
|
130 |
try:
|
131 |
batch_size = int(batch_size)
|
132 |
except Exception:
|
133 |
+
batch_size = 4
|
134 |
|
135 |
max_len = webui_config["tts_max"]
|
136 |
text = text.strip()[0:max_len]
|
|
|
172 |
prompt2=prompt2,
|
173 |
prefix=prefix,
|
174 |
)
|
|
|
|
|
175 |
else:
|
176 |
spliter = SentenceSplitter(webui_config["spliter_threshold"])
|
177 |
sentences = spliter.parse(text)
|
|
|
191 |
sample_rate = audio_data_batch[0][0]
|
192 |
audio_data = np.concatenate([data for _, data in audio_data_batch])
|
193 |
|
194 |
+
audio_data = audio.audio_to_int16(audio_data)
|
195 |
+
return sample_rate, audio_data
|
196 |
|
197 |
|
198 |
@torch.inference_mode()
|
|
|
380 |
batch_size_input = gr.Slider(
|
381 |
1,
|
382 |
webui_config["max_batch_size"],
|
383 |
+
value=4,
|
384 |
step=1,
|
385 |
label="Batch Size",
|
386 |
)
|
|
|
607 |
# batch size
|
608 |
batch_size_input = gr.Slider(
|
609 |
label="Batch Size",
|
610 |
+
value=4,
|
611 |
minimum=1,
|
612 |
maximum=webui_config["max_batch_size"],
|
613 |
step=1,
|
|
|
906 |
|
907 |
webui_config["tts_max"] = env.get_env_or_arg(args, "tts_max_len", 1000, int)
|
908 |
webui_config["ssml_max"] = env.get_env_or_arg(args, "ssml_max_len", 5000, int)
|
909 |
+
webui_config["max_batch_size"] = env.get_env_or_arg(args, "max_batch_size", 8, int)
|
910 |
|
911 |
demo = create_interface()
|
912 |
|