Spaces:
Running
Running
mrfakename
commited on
Commit
•
9c54d62
1
Parent(s):
d37849f
Sync from GitHub repo
Browse filesThis Space is synced from the GitHub repo: https://github.com/SWivid/F5-TTS. Please submit contributions to the Space there
- api.py +117 -0
- model/utils_infer.py +54 -20
api.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import soundfile as sf
|
2 |
+
import torch
|
3 |
+
import tqdm
|
4 |
+
from cached_path import cached_path
|
5 |
+
|
6 |
+
from model import DiT, UNetT
|
7 |
+
from model.utils import save_spectrogram
|
8 |
+
|
9 |
+
from model.utils_infer import load_vocoder, load_model, infer_process, remove_silence_for_generated_wav
|
10 |
+
|
11 |
+
|
12 |
+
class F5TTS:
|
13 |
+
def __init__(
|
14 |
+
self,
|
15 |
+
model_type="F5-TTS",
|
16 |
+
ckpt_file="",
|
17 |
+
vocab_file="",
|
18 |
+
ode_method="euler",
|
19 |
+
use_ema=True,
|
20 |
+
local_path=None,
|
21 |
+
device=None,
|
22 |
+
):
|
23 |
+
# Initialize parameters
|
24 |
+
self.final_wave = None
|
25 |
+
self.target_sample_rate = 24000
|
26 |
+
self.n_mel_channels = 100
|
27 |
+
self.hop_length = 256
|
28 |
+
self.target_rms = 0.1
|
29 |
+
|
30 |
+
# Set device
|
31 |
+
self.device = device or (
|
32 |
+
"cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
|
33 |
+
)
|
34 |
+
|
35 |
+
# Load models
|
36 |
+
self.load_vecoder_model(local_path)
|
37 |
+
self.load_ema_model(model_type, ckpt_file, vocab_file, ode_method, use_ema)
|
38 |
+
|
39 |
+
def load_vecoder_model(self, local_path):
|
40 |
+
self.vocos = load_vocoder(local_path is not None, local_path, self.device)
|
41 |
+
|
42 |
+
def load_ema_model(self, model_type, ckpt_file, vocab_file, ode_method, use_ema):
|
43 |
+
if model_type == "F5-TTS":
|
44 |
+
if not ckpt_file:
|
45 |
+
ckpt_file = str(cached_path("hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.safetensors"))
|
46 |
+
model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
|
47 |
+
model_cls = DiT
|
48 |
+
elif model_type == "E2-TTS":
|
49 |
+
if not ckpt_file:
|
50 |
+
ckpt_file = str(cached_path("hf://SWivid/E2-TTS/E2TTS_Base/model_1200000.safetensors"))
|
51 |
+
model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
52 |
+
model_cls = UNetT
|
53 |
+
else:
|
54 |
+
raise ValueError(f"Unknown model type: {model_type}")
|
55 |
+
|
56 |
+
self.ema_model = load_model(model_cls, model_cfg, ckpt_file, vocab_file, ode_method, use_ema, self.device)
|
57 |
+
|
58 |
+
def export_wav(self, wav, file_wave, remove_silence=False):
|
59 |
+
if remove_silence:
|
60 |
+
remove_silence_for_generated_wav(file_wave)
|
61 |
+
|
62 |
+
sf.write(file_wave, wav, self.target_sample_rate)
|
63 |
+
|
64 |
+
def export_spectrogram(self, spect, file_spect):
|
65 |
+
save_spectrogram(spect, file_spect)
|
66 |
+
|
67 |
+
def infer(
|
68 |
+
self,
|
69 |
+
ref_file,
|
70 |
+
ref_text,
|
71 |
+
gen_text,
|
72 |
+
sway_sampling_coef=-1,
|
73 |
+
cfg_strength=2,
|
74 |
+
nfe_step=32,
|
75 |
+
speed=1.0,
|
76 |
+
fix_duration=None,
|
77 |
+
remove_silence=False,
|
78 |
+
file_wave=None,
|
79 |
+
file_spect=None,
|
80 |
+
cross_fade_duration=0.15,
|
81 |
+
show_info=print,
|
82 |
+
progress=tqdm,
|
83 |
+
):
|
84 |
+
wav, sr, spect = infer_process(
|
85 |
+
ref_file,
|
86 |
+
ref_text,
|
87 |
+
gen_text,
|
88 |
+
self.ema_model,
|
89 |
+
cross_fade_duration,
|
90 |
+
speed,
|
91 |
+
show_info,
|
92 |
+
progress,
|
93 |
+
nfe_step,
|
94 |
+
cfg_strength,
|
95 |
+
sway_sampling_coef,
|
96 |
+
fix_duration,
|
97 |
+
)
|
98 |
+
|
99 |
+
if file_wave is not None:
|
100 |
+
self.export_wav(wav, file_wave, remove_silence)
|
101 |
+
|
102 |
+
if file_spect is not None:
|
103 |
+
self.export_spectrogram(spect, file_spect)
|
104 |
+
|
105 |
+
return wav, sr, spect
|
106 |
+
|
107 |
+
|
108 |
+
if __name__ == "__main__":
|
109 |
+
f5tts = F5TTS()
|
110 |
+
|
111 |
+
wav, sr, spect = f5tts.infer(
|
112 |
+
ref_file="tests/ref_audio/test_en_1_ref_short.wav",
|
113 |
+
ref_text="some call me nature, others call me mother nature.",
|
114 |
+
gen_text="""I don't really care what you call me. I've been a silent spectator, watching species evolve, empires rise and fall. But always remember, I am mighty and enduring. Respect me and I'll nurture you; ignore me and you shall face the consequences.""",
|
115 |
+
file_wave="tests/out.wav",
|
116 |
+
file_spect="tests/out.png",
|
117 |
+
)
|
model/utils_infer.py
CHANGED
@@ -38,12 +38,12 @@ target_sample_rate = 24000
|
|
38 |
n_mel_channels = 100
|
39 |
hop_length = 256
|
40 |
target_rms = 0.1
|
41 |
-
nfe_step = 32 # 16, 32
|
42 |
-
cfg_strength = 2.0
|
43 |
-
ode_method = "euler"
|
44 |
-
sway_sampling_coef = -1.0
|
45 |
-
speed = 1.0
|
46 |
-
fix_duration = None
|
47 |
|
48 |
# -----------------------------------------
|
49 |
|
@@ -84,7 +84,7 @@ def chunk_text(text, max_chars=135):
|
|
84 |
# load vocoder
|
85 |
|
86 |
|
87 |
-
def load_vocoder(is_local=False, local_path=""):
|
88 |
if is_local:
|
89 |
print(f"Load vocos from local path {local_path}")
|
90 |
vocos = Vocos.from_hparams(f"{local_path}/config.yaml")
|
@@ -100,14 +100,14 @@ def load_vocoder(is_local=False, local_path=""):
|
|
100 |
# load model for inference
|
101 |
|
102 |
|
103 |
-
def load_model(model_cls, model_cfg, ckpt_path, vocab_file=""):
|
104 |
if vocab_file == "":
|
105 |
vocab_file = "Emilia_ZH_EN"
|
106 |
tokenizer = "pinyin"
|
107 |
else:
|
108 |
tokenizer = "custom"
|
109 |
|
110 |
-
print("\nvocab : ", vocab_file
|
111 |
print("tokenizer : ", tokenizer)
|
112 |
print("model : ", ckpt_path, "\n")
|
113 |
|
@@ -125,7 +125,7 @@ def load_model(model_cls, model_cfg, ckpt_path, vocab_file=""):
|
|
125 |
vocab_char_map=vocab_char_map,
|
126 |
).to(device)
|
127 |
|
128 |
-
model = load_checkpoint(model, ckpt_path, device, use_ema=
|
129 |
|
130 |
return model
|
131 |
|
@@ -178,7 +178,18 @@ def preprocess_ref_audio_text(ref_audio_orig, ref_text, show_info=print):
|
|
178 |
|
179 |
|
180 |
def infer_process(
|
181 |
-
ref_audio,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
182 |
):
|
183 |
# Split the input text into batches
|
184 |
audio, sr = torchaudio.load(ref_audio)
|
@@ -188,14 +199,36 @@ def infer_process(
|
|
188 |
print(f"gen_text {i}", gen_text)
|
189 |
|
190 |
show_info(f"Generating audio in {len(gen_text_batches)} batches...")
|
191 |
-
return infer_batch_process(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
|
193 |
|
194 |
# infer batches
|
195 |
|
196 |
|
197 |
def infer_batch_process(
|
198 |
-
ref_audio,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
):
|
200 |
audio, sr = ref_audio
|
201 |
if audio.shape[0] > 1:
|
@@ -219,11 +252,14 @@ def infer_batch_process(
|
|
219 |
text_list = [ref_text + gen_text]
|
220 |
final_text_list = convert_char_to_pinyin(text_list)
|
221 |
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
|
|
|
|
|
|
227 |
|
228 |
# inference
|
229 |
with torch.inference_mode():
|
@@ -293,8 +329,6 @@ def infer_batch_process(
|
|
293 |
|
294 |
|
295 |
# remove silence from generated wav
|
296 |
-
|
297 |
-
|
298 |
def remove_silence_for_generated_wav(filename):
|
299 |
aseg = AudioSegment.from_file(filename)
|
300 |
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|
|
|
38 |
n_mel_channels = 100
|
39 |
hop_length = 256
|
40 |
target_rms = 0.1
|
41 |
+
# nfe_step = 32 # 16, 32
|
42 |
+
# cfg_strength = 2.0
|
43 |
+
# ode_method = "euler"
|
44 |
+
# sway_sampling_coef = -1.0
|
45 |
+
# speed = 1.0
|
46 |
+
# fix_duration = None
|
47 |
|
48 |
# -----------------------------------------
|
49 |
|
|
|
84 |
# load vocoder
|
85 |
|
86 |
|
87 |
+
def load_vocoder(is_local=False, local_path="", device=device):
|
88 |
if is_local:
|
89 |
print(f"Load vocos from local path {local_path}")
|
90 |
vocos = Vocos.from_hparams(f"{local_path}/config.yaml")
|
|
|
100 |
# load model for inference
|
101 |
|
102 |
|
103 |
+
def load_model(model_cls, model_cfg, ckpt_path, vocab_file="", ode_method="euler", use_ema=True, device=device):
|
104 |
if vocab_file == "":
|
105 |
vocab_file = "Emilia_ZH_EN"
|
106 |
tokenizer = "pinyin"
|
107 |
else:
|
108 |
tokenizer = "custom"
|
109 |
|
110 |
+
print("\nvocab : ", vocab_file)
|
111 |
print("tokenizer : ", tokenizer)
|
112 |
print("model : ", ckpt_path, "\n")
|
113 |
|
|
|
125 |
vocab_char_map=vocab_char_map,
|
126 |
).to(device)
|
127 |
|
128 |
+
model = load_checkpoint(model, ckpt_path, device, use_ema=use_ema)
|
129 |
|
130 |
return model
|
131 |
|
|
|
178 |
|
179 |
|
180 |
def infer_process(
|
181 |
+
ref_audio,
|
182 |
+
ref_text,
|
183 |
+
gen_text,
|
184 |
+
model_obj,
|
185 |
+
cross_fade_duration=0.15,
|
186 |
+
speed=1.0,
|
187 |
+
show_info=print,
|
188 |
+
progress=tqdm,
|
189 |
+
nfe_step=32,
|
190 |
+
cfg_strength=2,
|
191 |
+
sway_sampling_coef=-1,
|
192 |
+
fix_duration=None,
|
193 |
):
|
194 |
# Split the input text into batches
|
195 |
audio, sr = torchaudio.load(ref_audio)
|
|
|
199 |
print(f"gen_text {i}", gen_text)
|
200 |
|
201 |
show_info(f"Generating audio in {len(gen_text_batches)} batches...")
|
202 |
+
return infer_batch_process(
|
203 |
+
(audio, sr),
|
204 |
+
ref_text,
|
205 |
+
gen_text_batches,
|
206 |
+
model_obj,
|
207 |
+
cross_fade_duration,
|
208 |
+
speed,
|
209 |
+
progress,
|
210 |
+
nfe_step,
|
211 |
+
cfg_strength,
|
212 |
+
sway_sampling_coef,
|
213 |
+
fix_duration,
|
214 |
+
)
|
215 |
|
216 |
|
217 |
# infer batches
|
218 |
|
219 |
|
220 |
def infer_batch_process(
|
221 |
+
ref_audio,
|
222 |
+
ref_text,
|
223 |
+
gen_text_batches,
|
224 |
+
model_obj,
|
225 |
+
cross_fade_duration=0.15,
|
226 |
+
speed=1,
|
227 |
+
progress=tqdm,
|
228 |
+
nfe_step=32,
|
229 |
+
cfg_strength=2.0,
|
230 |
+
sway_sampling_coef=-1,
|
231 |
+
fix_duration=None,
|
232 |
):
|
233 |
audio, sr = ref_audio
|
234 |
if audio.shape[0] > 1:
|
|
|
252 |
text_list = [ref_text + gen_text]
|
253 |
final_text_list = convert_char_to_pinyin(text_list)
|
254 |
|
255 |
+
if fix_duration is not None:
|
256 |
+
duration = int(fix_duration * target_sample_rate / hop_length)
|
257 |
+
else:
|
258 |
+
# Calculate duration
|
259 |
+
ref_audio_len = audio.shape[-1] // hop_length
|
260 |
+
ref_text_len = len(ref_text.encode("utf-8"))
|
261 |
+
gen_text_len = len(gen_text.encode("utf-8"))
|
262 |
+
duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)
|
263 |
|
264 |
# inference
|
265 |
with torch.inference_mode():
|
|
|
329 |
|
330 |
|
331 |
# remove silence from generated wav
|
|
|
|
|
332 |
def remove_silence_for_generated_wav(filename):
|
333 |
aseg = AudioSegment.from_file(filename)
|
334 |
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|