Spaces:
Running
Running
mrfakename
commited on
Commit
•
d24a68b
1
Parent(s):
cea02d8
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
- app.py +35 -271
- inference-cli.py +33 -287
- model/utils_infer.py +306 -0
app.py
CHANGED
@@ -1,22 +1,25 @@
|
|
1 |
import re
|
2 |
-
import
|
3 |
-
|
|
|
4 |
import gradio as gr
|
5 |
import numpy as np
|
6 |
-
import
|
7 |
-
|
8 |
-
from pydub import AudioSegment, silence
|
9 |
-
from model import CFM, UNetT, DiT, MMDiT
|
10 |
from cached_path import cached_path
|
|
|
|
|
|
|
11 |
from model.utils import (
|
12 |
-
load_checkpoint,
|
13 |
-
get_tokenizer,
|
14 |
-
convert_char_to_pinyin,
|
15 |
save_spectrogram,
|
16 |
)
|
17 |
-
from
|
18 |
-
|
19 |
-
|
|
|
|
|
|
|
|
|
20 |
|
21 |
try:
|
22 |
import spaces
|
@@ -30,282 +33,47 @@ def gpu_decorator(func):
|
|
30 |
else:
|
31 |
return func
|
32 |
|
33 |
-
|
34 |
-
"cuda"
|
35 |
-
if torch.cuda.is_available()
|
36 |
-
else "mps" if torch.backends.mps.is_available() else "cpu"
|
37 |
-
)
|
38 |
-
|
39 |
-
print(f"Using {device} device")
|
40 |
-
|
41 |
-
pipe = pipeline(
|
42 |
-
"automatic-speech-recognition",
|
43 |
-
model="openai/whisper-large-v3-turbo",
|
44 |
-
torch_dtype=torch.float16,
|
45 |
-
device=device,
|
46 |
-
)
|
47 |
-
vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
|
48 |
-
|
49 |
-
# --------------------- Settings -------------------- #
|
50 |
-
|
51 |
-
target_sample_rate = 24000
|
52 |
-
n_mel_channels = 100
|
53 |
-
hop_length = 256
|
54 |
-
target_rms = 0.1
|
55 |
-
nfe_step = 32 # 16, 32
|
56 |
-
cfg_strength = 2.0
|
57 |
-
ode_method = "euler"
|
58 |
-
sway_sampling_coef = -1.0
|
59 |
-
speed = 1.0
|
60 |
-
fix_duration = None
|
61 |
-
|
62 |
-
|
63 |
-
def load_model(repo_name, exp_name, model_cls, model_cfg, ckpt_step):
|
64 |
-
ckpt_path = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
|
65 |
-
# ckpt_path = f"ckpts/{exp_name}/model_{ckpt_step}.pt" # .pt | .safetensors
|
66 |
-
vocab_char_map, vocab_size = get_tokenizer("Emilia_ZH_EN", "pinyin")
|
67 |
-
model = CFM(
|
68 |
-
transformer=model_cls(
|
69 |
-
**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels
|
70 |
-
),
|
71 |
-
mel_spec_kwargs=dict(
|
72 |
-
target_sample_rate=target_sample_rate,
|
73 |
-
n_mel_channels=n_mel_channels,
|
74 |
-
hop_length=hop_length,
|
75 |
-
),
|
76 |
-
odeint_kwargs=dict(
|
77 |
-
method=ode_method,
|
78 |
-
),
|
79 |
-
vocab_char_map=vocab_char_map,
|
80 |
-
).to(device)
|
81 |
-
|
82 |
-
model = load_checkpoint(model, ckpt_path, device, use_ema = True)
|
83 |
-
|
84 |
-
return model
|
85 |
|
86 |
|
87 |
# load models
|
88 |
-
F5TTS_model_cfg = dict(
|
89 |
-
|
90 |
-
)
|
91 |
-
E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
92 |
-
|
93 |
-
F5TTS_ema_model = load_model(
|
94 |
-
"F5-TTS", "F5TTS_Base", DiT, F5TTS_model_cfg, 1200000
|
95 |
-
)
|
96 |
-
E2TTS_ema_model = load_model(
|
97 |
-
"E2-TTS", "E2TTS_Base", UNetT, E2TTS_model_cfg, 1200000
|
98 |
-
)
|
99 |
-
|
100 |
-
def chunk_text(text, max_chars=135):
|
101 |
-
"""
|
102 |
-
Splits the input text into chunks, each with a maximum number of characters.
|
103 |
-
|
104 |
-
Args:
|
105 |
-
text (str): The text to be split.
|
106 |
-
max_chars (int): The maximum number of characters per chunk.
|
107 |
-
|
108 |
-
Returns:
|
109 |
-
List[str]: A list of text chunks.
|
110 |
-
"""
|
111 |
-
chunks = []
|
112 |
-
current_chunk = ""
|
113 |
-
# Split the text into sentences based on punctuation followed by whitespace
|
114 |
-
sentences = re.split(r'(?<=[;:,.!?])\s+|(?<=[;:,。!?])', text)
|
115 |
-
|
116 |
-
for sentence in sentences:
|
117 |
-
if len(current_chunk.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
|
118 |
-
current_chunk += sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
|
119 |
-
else:
|
120 |
-
if current_chunk:
|
121 |
-
chunks.append(current_chunk.strip())
|
122 |
-
current_chunk = sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
|
123 |
|
124 |
-
|
125 |
-
|
126 |
|
127 |
-
return chunks
|
128 |
|
129 |
@gpu_decorator
|
130 |
-
def
|
131 |
-
|
|
|
|
|
|
|
132 |
ema_model = F5TTS_ema_model
|
133 |
-
elif
|
134 |
ema_model = E2TTS_ema_model
|
135 |
|
136 |
-
|
137 |
-
if audio.shape[0] > 1:
|
138 |
-
audio = torch.mean(audio, dim=0, keepdim=True)
|
139 |
-
|
140 |
-
rms = torch.sqrt(torch.mean(torch.square(audio)))
|
141 |
-
if rms < target_rms:
|
142 |
-
audio = audio * target_rms / rms
|
143 |
-
if sr != target_sample_rate:
|
144 |
-
resampler = torchaudio.transforms.Resample(sr, target_sample_rate)
|
145 |
-
audio = resampler(audio)
|
146 |
-
audio = audio.to(device)
|
147 |
-
|
148 |
-
generated_waves = []
|
149 |
-
spectrograms = []
|
150 |
-
|
151 |
-
if len(ref_text[-1].encode('utf-8')) == 1:
|
152 |
-
ref_text = ref_text + " "
|
153 |
-
for i, gen_text in enumerate(progress.tqdm(gen_text_batches)):
|
154 |
-
# Prepare the text
|
155 |
-
text_list = [ref_text + gen_text]
|
156 |
-
final_text_list = convert_char_to_pinyin(text_list)
|
157 |
-
|
158 |
-
# Calculate duration
|
159 |
-
ref_audio_len = audio.shape[-1] // hop_length
|
160 |
-
ref_text_len = len(ref_text.encode('utf-8'))
|
161 |
-
gen_text_len = len(gen_text.encode('utf-8'))
|
162 |
-
duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)
|
163 |
-
|
164 |
-
# inference
|
165 |
-
with torch.inference_mode():
|
166 |
-
generated, _ = ema_model.sample(
|
167 |
-
cond=audio,
|
168 |
-
text=final_text_list,
|
169 |
-
duration=duration,
|
170 |
-
steps=nfe_step,
|
171 |
-
cfg_strength=cfg_strength,
|
172 |
-
sway_sampling_coef=sway_sampling_coef,
|
173 |
-
)
|
174 |
-
|
175 |
-
generated = generated.to(torch.float32)
|
176 |
-
generated = generated[:, ref_audio_len:, :]
|
177 |
-
generated_mel_spec = generated.permute(0, 2, 1)
|
178 |
-
generated_wave = vocos.decode(generated_mel_spec.cpu())
|
179 |
-
if rms < target_rms:
|
180 |
-
generated_wave = generated_wave * rms / target_rms
|
181 |
-
|
182 |
-
# wav -> numpy
|
183 |
-
generated_wave = generated_wave.squeeze().cpu().numpy()
|
184 |
-
|
185 |
-
generated_waves.append(generated_wave)
|
186 |
-
spectrograms.append(generated_mel_spec[0].cpu().numpy())
|
187 |
-
|
188 |
-
# Combine all generated waves with cross-fading
|
189 |
-
if cross_fade_duration <= 0:
|
190 |
-
# Simply concatenate
|
191 |
-
final_wave = np.concatenate(generated_waves)
|
192 |
-
else:
|
193 |
-
final_wave = generated_waves[0]
|
194 |
-
for i in range(1, len(generated_waves)):
|
195 |
-
prev_wave = final_wave
|
196 |
-
next_wave = generated_waves[i]
|
197 |
-
|
198 |
-
# Calculate cross-fade samples, ensuring it does not exceed wave lengths
|
199 |
-
cross_fade_samples = int(cross_fade_duration * target_sample_rate)
|
200 |
-
cross_fade_samples = min(cross_fade_samples, len(prev_wave), len(next_wave))
|
201 |
-
|
202 |
-
if cross_fade_samples <= 0:
|
203 |
-
# No overlap possible, concatenate
|
204 |
-
final_wave = np.concatenate([prev_wave, next_wave])
|
205 |
-
continue
|
206 |
-
|
207 |
-
# Overlapping parts
|
208 |
-
prev_overlap = prev_wave[-cross_fade_samples:]
|
209 |
-
next_overlap = next_wave[:cross_fade_samples]
|
210 |
-
|
211 |
-
# Fade out and fade in
|
212 |
-
fade_out = np.linspace(1, 0, cross_fade_samples)
|
213 |
-
fade_in = np.linspace(0, 1, cross_fade_samples)
|
214 |
-
|
215 |
-
# Cross-faded overlap
|
216 |
-
cross_faded_overlap = prev_overlap * fade_out + next_overlap * fade_in
|
217 |
-
|
218 |
-
# Combine
|
219 |
-
new_wave = np.concatenate([
|
220 |
-
prev_wave[:-cross_fade_samples],
|
221 |
-
cross_faded_overlap,
|
222 |
-
next_wave[cross_fade_samples:]
|
223 |
-
])
|
224 |
-
|
225 |
-
final_wave = new_wave
|
226 |
|
227 |
# Remove silence
|
228 |
if remove_silence:
|
229 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
230 |
-
sf.write(f.name, final_wave,
|
231 |
-
|
232 |
-
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|
233 |
-
non_silent_wave = AudioSegment.silent(duration=0)
|
234 |
-
for non_silent_seg in non_silent_segs:
|
235 |
-
non_silent_wave += non_silent_seg
|
236 |
-
aseg = non_silent_wave
|
237 |
-
aseg.export(f.name, format="wav")
|
238 |
final_wave, _ = torchaudio.load(f.name)
|
239 |
final_wave = final_wave.squeeze().cpu().numpy()
|
240 |
|
241 |
-
#
|
242 |
-
combined_spectrogram = np.concatenate(spectrograms, axis=1)
|
243 |
-
|
244 |
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_spectrogram:
|
245 |
spectrogram_path = tmp_spectrogram.name
|
246 |
save_spectrogram(combined_spectrogram, spectrogram_path)
|
247 |
|
248 |
-
return (
|
249 |
-
|
250 |
-
@gpu_decorator
|
251 |
-
def infer(ref_audio_orig, ref_text, gen_text, exp_name, remove_silence, cross_fade_duration=0.15):
|
252 |
-
|
253 |
-
print(gen_text)
|
254 |
-
|
255 |
-
gr.Info("Converting audio...")
|
256 |
-
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
257 |
-
aseg = AudioSegment.from_file(ref_audio_orig)
|
258 |
-
|
259 |
-
non_silent_segs = silence.split_on_silence(
|
260 |
-
aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=1000
|
261 |
-
)
|
262 |
-
non_silent_wave = AudioSegment.silent(duration=0)
|
263 |
-
for non_silent_seg in non_silent_segs:
|
264 |
-
non_silent_wave += non_silent_seg
|
265 |
-
aseg = non_silent_wave
|
266 |
-
|
267 |
-
audio_duration = len(aseg)
|
268 |
-
if audio_duration > 15000:
|
269 |
-
gr.Warning("Audio is over 15s, clipping to only first 15s.")
|
270 |
-
aseg = aseg[:15000]
|
271 |
-
aseg.export(f.name, format="wav")
|
272 |
-
ref_audio = f.name
|
273 |
-
|
274 |
-
if not ref_text.strip():
|
275 |
-
gr.Info("No reference text provided, transcribing reference audio...")
|
276 |
-
ref_text = pipe(
|
277 |
-
ref_audio,
|
278 |
-
chunk_length_s=30,
|
279 |
-
batch_size=128,
|
280 |
-
generate_kwargs={"task": "transcribe"},
|
281 |
-
return_timestamps=False,
|
282 |
-
)["text"].strip()
|
283 |
-
gr.Info("Finished transcription")
|
284 |
-
else:
|
285 |
-
gr.Info("Using custom reference text...")
|
286 |
-
|
287 |
-
# Add the functionality to ensure it ends with ". "
|
288 |
-
if not ref_text.endswith(". "):
|
289 |
-
if ref_text.endswith("."):
|
290 |
-
ref_text += " "
|
291 |
-
else:
|
292 |
-
ref_text += ". "
|
293 |
-
|
294 |
-
audio, sr = torchaudio.load(ref_audio)
|
295 |
-
|
296 |
-
# Use the new chunk_text function to split gen_text
|
297 |
-
max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (25 - audio.shape[-1] / sr))
|
298 |
-
gen_text_batches = chunk_text(gen_text, max_chars=max_chars)
|
299 |
-
print('ref_text', ref_text)
|
300 |
-
for i, batch_text in enumerate(gen_text_batches):
|
301 |
-
print(f'gen_text {i}', batch_text)
|
302 |
-
|
303 |
-
gr.Info(f"Generating audio using {exp_name} in {len(gen_text_batches)} batches")
|
304 |
-
return infer_batch((audio, sr), ref_text, gen_text_batches, exp_name, remove_silence, cross_fade_duration)
|
305 |
|
306 |
|
307 |
@gpu_decorator
|
308 |
-
def generate_podcast(script, speaker1_name, ref_audio1, ref_text1, speaker2_name, ref_audio2, ref_text2,
|
309 |
# Split the script into speaker blocks
|
310 |
speaker_pattern = re.compile(f"^({re.escape(speaker1_name)}|{re.escape(speaker2_name)}):", re.MULTILINE)
|
311 |
speaker_blocks = speaker_pattern.split(script)[1:] # Skip the first empty element
|
@@ -327,7 +95,7 @@ def generate_podcast(script, speaker1_name, ref_audio1, ref_text1, speaker2_name
|
|
327 |
continue # Skip if the speaker is neither speaker1 nor speaker2
|
328 |
|
329 |
# Generate audio for this block
|
330 |
-
audio, _ = infer(ref_audio, ref_text, text,
|
331 |
|
332 |
# Convert the generated audio to a numpy array
|
333 |
sr, audio_data = audio
|
@@ -377,10 +145,6 @@ def parse_speechtypes_text(gen_text):
|
|
377 |
|
378 |
return segments
|
379 |
|
380 |
-
def update_speed(new_speed):
|
381 |
-
global speed
|
382 |
-
speed = new_speed
|
383 |
-
return f"Speed set to: {speed}"
|
384 |
|
385 |
with gr.Blocks() as app_credits:
|
386 |
gr.Markdown("""
|
@@ -413,7 +177,7 @@ with gr.Blocks() as app_tts:
|
|
413 |
label="Speed",
|
414 |
minimum=0.3,
|
415 |
maximum=2.0,
|
416 |
-
value=
|
417 |
step=0.1,
|
418 |
info="Adjust the speed of the audio.",
|
419 |
)
|
@@ -425,7 +189,6 @@ with gr.Blocks() as app_tts:
|
|
425 |
step=0.01,
|
426 |
info="Set the duration of the cross-fade between audio clips.",
|
427 |
)
|
428 |
-
speed_slider.change(update_speed, inputs=speed_slider)
|
429 |
|
430 |
audio_output = gr.Audio(label="Synthesized Audio")
|
431 |
spectrogram_output = gr.Image(label="Spectrogram")
|
@@ -439,6 +202,7 @@ with gr.Blocks() as app_tts:
|
|
439 |
model_choice,
|
440 |
remove_silence,
|
441 |
cross_fade_duration_slider,
|
|
|
442 |
],
|
443 |
outputs=[audio_output, spectrogram_output],
|
444 |
)
|
|
|
1 |
import re
|
2 |
+
import tempfile
|
3 |
+
|
4 |
+
import click
|
5 |
import gradio as gr
|
6 |
import numpy as np
|
7 |
+
import soundfile as sf
|
8 |
+
import torchaudio
|
|
|
|
|
9 |
from cached_path import cached_path
|
10 |
+
from pydub import AudioSegment
|
11 |
+
|
12 |
+
from model import DiT, UNetT
|
13 |
from model.utils import (
|
|
|
|
|
|
|
14 |
save_spectrogram,
|
15 |
)
|
16 |
+
from model.utils_infer import (
|
17 |
+
load_vocoder,
|
18 |
+
load_model,
|
19 |
+
preprocess_ref_audio_text,
|
20 |
+
infer_process,
|
21 |
+
remove_silence_for_generated_wav,
|
22 |
+
)
|
23 |
|
24 |
try:
|
25 |
import spaces
|
|
|
33 |
else:
|
34 |
return func
|
35 |
|
36 |
+
vocos = load_vocoder()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
|
39 |
# load models
|
40 |
+
F5TTS_model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
|
41 |
+
F5TTS_ema_model = load_model(DiT, F5TTS_model_cfg, str(cached_path(f"hf://SWivid/F5-TTS/F5TTS_Base/model_1200000.safetensors")))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
+
E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
44 |
+
E2TTS_ema_model = load_model(UNetT, E2TTS_model_cfg, str(cached_path(f"hf://SWivid/E2-TTS/E2TTS_Base/model_1200000.safetensors")))
|
45 |
|
|
|
46 |
|
47 |
@gpu_decorator
|
48 |
+
def infer(ref_audio_orig, ref_text, gen_text, model, remove_silence, cross_fade_duration=0.15, speed=1):
|
49 |
+
|
50 |
+
ref_audio, ref_text = preprocess_ref_audio_text(ref_audio_orig, ref_text, show_info=gr.Info)
|
51 |
+
|
52 |
+
if model == "F5-TTS":
|
53 |
ema_model = F5TTS_ema_model
|
54 |
+
elif model == "E2-TTS":
|
55 |
ema_model = E2TTS_ema_model
|
56 |
|
57 |
+
final_wave, final_sample_rate, combined_spectrogram = infer_process(ref_audio, ref_text, gen_text, ema_model, cross_fade_duration=cross_fade_duration, speed=speed, show_info=gr.Info, progress=gr.Progress())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
59 |
# Remove silence
|
60 |
if remove_silence:
|
61 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
62 |
+
sf.write(f.name, final_wave, final_sample_rate)
|
63 |
+
remove_silence_for_generated_wav(f.name)
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
final_wave, _ = torchaudio.load(f.name)
|
65 |
final_wave = final_wave.squeeze().cpu().numpy()
|
66 |
|
67 |
+
# Save the spectrogram
|
|
|
|
|
68 |
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_spectrogram:
|
69 |
spectrogram_path = tmp_spectrogram.name
|
70 |
save_spectrogram(combined_spectrogram, spectrogram_path)
|
71 |
|
72 |
+
return (final_sample_rate, final_wave), spectrogram_path
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
|
75 |
@gpu_decorator
|
76 |
+
def generate_podcast(script, speaker1_name, ref_audio1, ref_text1, speaker2_name, ref_audio2, ref_text2, model, remove_silence):
|
77 |
# Split the script into speaker blocks
|
78 |
speaker_pattern = re.compile(f"^({re.escape(speaker1_name)}|{re.escape(speaker2_name)}):", re.MULTILINE)
|
79 |
speaker_blocks = speaker_pattern.split(script)[1:] # Skip the first empty element
|
|
|
95 |
continue # Skip if the speaker is neither speaker1 nor speaker2
|
96 |
|
97 |
# Generate audio for this block
|
98 |
+
audio, _ = infer(ref_audio, ref_text, text, model, remove_silence)
|
99 |
|
100 |
# Convert the generated audio to a numpy array
|
101 |
sr, audio_data = audio
|
|
|
145 |
|
146 |
return segments
|
147 |
|
|
|
|
|
|
|
|
|
148 |
|
149 |
with gr.Blocks() as app_credits:
|
150 |
gr.Markdown("""
|
|
|
177 |
label="Speed",
|
178 |
minimum=0.3,
|
179 |
maximum=2.0,
|
180 |
+
value=1.0,
|
181 |
step=0.1,
|
182 |
info="Adjust the speed of the audio.",
|
183 |
)
|
|
|
189 |
step=0.01,
|
190 |
info="Set the duration of the cross-fade between audio clips.",
|
191 |
)
|
|
|
192 |
|
193 |
audio_output = gr.Audio(label="Synthesized Audio")
|
194 |
spectrogram_output = gr.Image(label="Spectrogram")
|
|
|
202 |
model_choice,
|
203 |
remove_silence,
|
204 |
cross_fade_duration_slider,
|
205 |
+
speed_slider,
|
206 |
],
|
207 |
outputs=[audio_output, spectrogram_output],
|
208 |
)
|
inference-cli.py
CHANGED
@@ -1,23 +1,22 @@
|
|
1 |
import argparse
|
2 |
import codecs
|
3 |
import re
|
4 |
-
import tempfile
|
5 |
from pathlib import Path
|
6 |
|
7 |
import numpy as np
|
8 |
import soundfile as sf
|
9 |
import tomli
|
10 |
-
import torch
|
11 |
-
import torchaudio
|
12 |
-
import tqdm
|
13 |
from cached_path import cached_path
|
14 |
-
from pydub import AudioSegment, silence
|
15 |
-
from transformers import pipeline
|
16 |
-
from vocos import Vocos
|
17 |
|
18 |
-
from model import
|
19 |
-
from model.
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
parser = argparse.ArgumentParser(
|
23 |
prog="python3 inference-cli.py",
|
@@ -104,282 +103,35 @@ wave_path = Path(output_dir)/"out.wav"
|
|
104 |
spectrogram_path = Path(output_dir)/"out.png"
|
105 |
vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
|
106 |
|
107 |
-
|
108 |
-
"cuda"
|
109 |
-
if torch.cuda.is_available()
|
110 |
-
else "mps" if torch.backends.mps.is_available() else "cpu"
|
111 |
-
)
|
112 |
-
|
113 |
-
if args.load_vocoder_from_local:
|
114 |
-
print(f"Load vocos from local path {vocos_local_path}")
|
115 |
-
vocos = Vocos.from_hparams(f"{vocos_local_path}/config.yaml")
|
116 |
-
state_dict = torch.load(f"{vocos_local_path}/pytorch_model.bin", map_location=device)
|
117 |
-
vocos.load_state_dict(state_dict)
|
118 |
-
vocos.eval()
|
119 |
-
else:
|
120 |
-
print("Download Vocos from huggingface charactr/vocos-mel-24khz")
|
121 |
-
vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
|
122 |
-
|
123 |
-
print(f"Using {device} device")
|
124 |
-
|
125 |
-
# --------------------- Settings -------------------- #
|
126 |
-
|
127 |
-
target_sample_rate = 24000
|
128 |
-
n_mel_channels = 100
|
129 |
-
hop_length = 256
|
130 |
-
target_rms = 0.1
|
131 |
-
nfe_step = 32 # 16, 32
|
132 |
-
cfg_strength = 2.0
|
133 |
-
ode_method = "euler"
|
134 |
-
sway_sampling_coef = -1.0
|
135 |
-
speed = 1.0
|
136 |
-
# fix_duration = 27 # None or float (duration in seconds)
|
137 |
-
fix_duration = None
|
138 |
|
139 |
-
def load_model(model_cls, model_cfg, ckpt_path,file_vocab):
|
140 |
-
|
141 |
-
if file_vocab=="":
|
142 |
-
file_vocab="Emilia_ZH_EN"
|
143 |
-
tokenizer="pinyin"
|
144 |
-
else:
|
145 |
-
tokenizer="custom"
|
146 |
-
|
147 |
-
print("\nvocab : ", vocab_file,tokenizer)
|
148 |
-
print("tokenizer : ", tokenizer)
|
149 |
-
print("model : ", ckpt_path,"\n")
|
150 |
-
|
151 |
-
vocab_char_map, vocab_size = get_tokenizer(file_vocab, tokenizer)
|
152 |
-
model = CFM(
|
153 |
-
transformer=model_cls(
|
154 |
-
**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels
|
155 |
-
),
|
156 |
-
mel_spec_kwargs=dict(
|
157 |
-
target_sample_rate=target_sample_rate,
|
158 |
-
n_mel_channels=n_mel_channels,
|
159 |
-
hop_length=hop_length,
|
160 |
-
),
|
161 |
-
odeint_kwargs=dict(
|
162 |
-
method=ode_method,
|
163 |
-
),
|
164 |
-
vocab_char_map=vocab_char_map,
|
165 |
-
).to(device)
|
166 |
-
|
167 |
-
model = load_checkpoint(model, ckpt_path, device, use_ema = True)
|
168 |
-
|
169 |
-
return model
|
170 |
|
171 |
# load models
|
172 |
-
F5TTS_model_cfg = dict(
|
173 |
-
dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4
|
174 |
-
)
|
175 |
-
E2TTS_model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
176 |
-
|
177 |
if model == "F5-TTS":
|
178 |
-
|
|
|
179 |
if ckpt_file == "":
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
ema_model = load_model(DiT, F5TTS_model_cfg, ckpt_file,vocab_file)
|
186 |
|
187 |
elif model == "E2-TTS":
|
|
|
|
|
188 |
if ckpt_file == "":
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
"automatic-speech-recognition",
|
198 |
-
model="openai/whisper-large-v3-turbo",
|
199 |
-
torch_dtype=torch.float16,
|
200 |
-
device=device,
|
201 |
-
)
|
202 |
-
|
203 |
-
def chunk_text(text, max_chars=135):
|
204 |
-
"""
|
205 |
-
Splits the input text into chunks, each with a maximum number of characters.
|
206 |
-
Args:
|
207 |
-
text (str): The text to be split.
|
208 |
-
max_chars (int): The maximum number of characters per chunk.
|
209 |
-
Returns:
|
210 |
-
List[str]: A list of text chunks.
|
211 |
-
"""
|
212 |
-
chunks = []
|
213 |
-
current_chunk = ""
|
214 |
-
# Split the text into sentences based on punctuation followed by whitespace
|
215 |
-
sentences = re.split(r'(?<=[;:,.!?])\s+|(?<=[;:,。!?])', text)
|
216 |
-
|
217 |
-
for sentence in sentences:
|
218 |
-
if len(current_chunk.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
|
219 |
-
current_chunk += sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
|
220 |
-
else:
|
221 |
-
if current_chunk:
|
222 |
-
chunks.append(current_chunk.strip())
|
223 |
-
current_chunk = sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
|
224 |
-
|
225 |
-
if current_chunk:
|
226 |
-
chunks.append(current_chunk.strip())
|
227 |
-
|
228 |
-
return chunks
|
229 |
-
|
230 |
-
#ckpt_path = f"ckpts/{exp_name}/model_{ckpt_step}.pt" # .pt | .safetensors
|
231 |
-
#if not Path(ckpt_path).exists():
|
232 |
-
#ckpt_path = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
|
233 |
-
|
234 |
-
def infer_batch(ref_audio, ref_text, gen_text_batches, model, remove_silence, cross_fade_duration=0.15):
|
235 |
-
audio, sr = ref_audio
|
236 |
-
if audio.shape[0] > 1:
|
237 |
-
audio = torch.mean(audio, dim=0, keepdim=True)
|
238 |
-
|
239 |
-
rms = torch.sqrt(torch.mean(torch.square(audio)))
|
240 |
-
if rms < target_rms:
|
241 |
-
audio = audio * target_rms / rms
|
242 |
-
if sr != target_sample_rate:
|
243 |
-
resampler = torchaudio.transforms.Resample(sr, target_sample_rate)
|
244 |
-
audio = resampler(audio)
|
245 |
-
audio = audio.to(device)
|
246 |
-
|
247 |
-
generated_waves = []
|
248 |
-
spectrograms = []
|
249 |
-
|
250 |
-
if len(ref_text[-1].encode('utf-8')) == 1:
|
251 |
-
ref_text = ref_text + " "
|
252 |
-
for i, gen_text in enumerate(tqdm.tqdm(gen_text_batches)):
|
253 |
-
# Prepare the text
|
254 |
-
text_list = [ref_text + gen_text]
|
255 |
-
final_text_list = convert_char_to_pinyin(text_list)
|
256 |
-
|
257 |
-
# Calculate duration
|
258 |
-
ref_audio_len = audio.shape[-1] // hop_length
|
259 |
-
ref_text_len = len(ref_text.encode('utf-8'))
|
260 |
-
gen_text_len = len(gen_text.encode('utf-8'))
|
261 |
-
duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)
|
262 |
-
|
263 |
-
# inference
|
264 |
-
with torch.inference_mode():
|
265 |
-
generated, _ = ema_model.sample(
|
266 |
-
cond=audio,
|
267 |
-
text=final_text_list,
|
268 |
-
duration=duration,
|
269 |
-
steps=nfe_step,
|
270 |
-
cfg_strength=cfg_strength,
|
271 |
-
sway_sampling_coef=sway_sampling_coef,
|
272 |
-
)
|
273 |
-
|
274 |
-
generated = generated.to(torch.float32)
|
275 |
-
generated = generated[:, ref_audio_len:, :]
|
276 |
-
generated_mel_spec = generated.permute(0, 2, 1)
|
277 |
-
generated_wave = vocos.decode(generated_mel_spec.cpu())
|
278 |
-
if rms < target_rms:
|
279 |
-
generated_wave = generated_wave * rms / target_rms
|
280 |
-
|
281 |
-
# wav -> numpy
|
282 |
-
generated_wave = generated_wave.squeeze().cpu().numpy()
|
283 |
-
|
284 |
-
generated_waves.append(generated_wave)
|
285 |
-
spectrograms.append(generated_mel_spec[0].cpu().numpy())
|
286 |
-
|
287 |
-
# Combine all generated waves with cross-fading
|
288 |
-
if cross_fade_duration <= 0:
|
289 |
-
# Simply concatenate
|
290 |
-
final_wave = np.concatenate(generated_waves)
|
291 |
-
else:
|
292 |
-
final_wave = generated_waves[0]
|
293 |
-
for i in range(1, len(generated_waves)):
|
294 |
-
prev_wave = final_wave
|
295 |
-
next_wave = generated_waves[i]
|
296 |
-
|
297 |
-
# Calculate cross-fade samples, ensuring it does not exceed wave lengths
|
298 |
-
cross_fade_samples = int(cross_fade_duration * target_sample_rate)
|
299 |
-
cross_fade_samples = min(cross_fade_samples, len(prev_wave), len(next_wave))
|
300 |
-
|
301 |
-
if cross_fade_samples <= 0:
|
302 |
-
# No overlap possible, concatenate
|
303 |
-
final_wave = np.concatenate([prev_wave, next_wave])
|
304 |
-
continue
|
305 |
-
|
306 |
-
# Overlapping parts
|
307 |
-
prev_overlap = prev_wave[-cross_fade_samples:]
|
308 |
-
next_overlap = next_wave[:cross_fade_samples]
|
309 |
-
|
310 |
-
# Fade out and fade in
|
311 |
-
fade_out = np.linspace(1, 0, cross_fade_samples)
|
312 |
-
fade_in = np.linspace(0, 1, cross_fade_samples)
|
313 |
-
|
314 |
-
# Cross-faded overlap
|
315 |
-
cross_faded_overlap = prev_overlap * fade_out + next_overlap * fade_in
|
316 |
-
|
317 |
-
# Combine
|
318 |
-
new_wave = np.concatenate([
|
319 |
-
prev_wave[:-cross_fade_samples],
|
320 |
-
cross_faded_overlap,
|
321 |
-
next_wave[cross_fade_samples:]
|
322 |
-
])
|
323 |
-
|
324 |
-
final_wave = new_wave
|
325 |
-
|
326 |
-
# Create a combined spectrogram
|
327 |
-
combined_spectrogram = np.concatenate(spectrograms, axis=1)
|
328 |
-
|
329 |
-
return final_wave, combined_spectrogram
|
330 |
-
|
331 |
-
def process_voice(ref_audio_orig, ref_text):
|
332 |
-
print("Converting", ref_audio_orig)
|
333 |
-
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
334 |
-
aseg = AudioSegment.from_file(ref_audio_orig)
|
335 |
-
|
336 |
-
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=1000)
|
337 |
-
non_silent_wave = AudioSegment.silent(duration=0)
|
338 |
-
for non_silent_seg in non_silent_segs:
|
339 |
-
non_silent_wave += non_silent_seg
|
340 |
-
aseg = non_silent_wave
|
341 |
-
|
342 |
-
audio_duration = len(aseg)
|
343 |
-
if audio_duration > 15000:
|
344 |
-
print("Audio is over 15s, clipping to only first 15s.")
|
345 |
-
aseg = aseg[:15000]
|
346 |
-
aseg.export(f.name, format="wav")
|
347 |
-
ref_audio = f.name
|
348 |
-
|
349 |
-
if not ref_text.strip():
|
350 |
-
print("No reference text provided, transcribing reference audio...")
|
351 |
-
ref_text = asr_pipe(
|
352 |
-
ref_audio,
|
353 |
-
chunk_length_s=30,
|
354 |
-
batch_size=128,
|
355 |
-
generate_kwargs={"task": "transcribe"},
|
356 |
-
return_timestamps=False,
|
357 |
-
)["text"].strip()
|
358 |
-
print("Finished transcription")
|
359 |
-
else:
|
360 |
-
print("Using custom reference text...")
|
361 |
-
return ref_audio, ref_text
|
362 |
-
|
363 |
-
def infer(ref_audio, ref_text, gen_text, model, remove_silence, cross_fade_duration=0.15):
|
364 |
-
# Add the functionality to ensure it ends with ". "
|
365 |
-
if not ref_text.endswith(". ") and not ref_text.endswith("。"):
|
366 |
-
if ref_text.endswith("."):
|
367 |
-
ref_text += " "
|
368 |
-
else:
|
369 |
-
ref_text += ". "
|
370 |
-
|
371 |
-
# Split the input text into batches
|
372 |
-
audio, sr = torchaudio.load(ref_audio)
|
373 |
-
max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (25 - audio.shape[-1] / sr))
|
374 |
-
gen_text_batches = chunk_text(gen_text, max_chars=max_chars)
|
375 |
-
for i, gen_text in enumerate(gen_text_batches):
|
376 |
-
print(f'gen_text {i}', gen_text)
|
377 |
-
|
378 |
-
print(f"Generating audio using {model} in {len(gen_text_batches)} batches, loading models...")
|
379 |
-
return infer_batch((audio, sr), ref_text, gen_text_batches, model, remove_silence, cross_fade_duration)
|
380 |
|
381 |
|
382 |
-
def
|
383 |
main_voice = {"ref_audio":ref_audio, "ref_text":ref_text}
|
384 |
if "voices" not in config:
|
385 |
voices = {"main": main_voice}
|
@@ -387,7 +139,7 @@ def process(ref_audio, ref_text, text_gen, model, remove_silence):
|
|
387 |
voices = config["voices"]
|
388 |
voices["main"] = main_voice
|
389 |
for voice in voices:
|
390 |
-
voices[voice]['ref_audio'], voices[voice]['ref_text'] =
|
391 |
print("Voice:", voice)
|
392 |
print("Ref_audio:", voices[voice]['ref_audio'])
|
393 |
print("Ref_text:", voices[voice]['ref_text'])
|
@@ -407,23 +159,17 @@ def process(ref_audio, ref_text, text_gen, model, remove_silence):
|
|
407 |
ref_audio = voices[voice]['ref_audio']
|
408 |
ref_text = voices[voice]['ref_text']
|
409 |
print(f"Voice: {voice}")
|
410 |
-
audio, spectragram =
|
411 |
generated_audio_segments.append(audio)
|
412 |
|
413 |
if generated_audio_segments:
|
414 |
final_wave = np.concatenate(generated_audio_segments)
|
415 |
with open(wave_path, "wb") as f:
|
416 |
-
sf.write(f.name, final_wave,
|
417 |
# Remove silence
|
418 |
if remove_silence:
|
419 |
-
|
420 |
-
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|
421 |
-
non_silent_wave = AudioSegment.silent(duration=0)
|
422 |
-
for non_silent_seg in non_silent_segs:
|
423 |
-
non_silent_wave += non_silent_seg
|
424 |
-
aseg = non_silent_wave
|
425 |
-
aseg.export(f.name, format="wav")
|
426 |
print(f.name)
|
427 |
|
428 |
|
429 |
-
|
|
|
1 |
import argparse
|
2 |
import codecs
|
3 |
import re
|
|
|
4 |
from pathlib import Path
|
5 |
|
6 |
import numpy as np
|
7 |
import soundfile as sf
|
8 |
import tomli
|
|
|
|
|
|
|
9 |
from cached_path import cached_path
|
|
|
|
|
|
|
10 |
|
11 |
+
from model import DiT, UNetT
|
12 |
+
from model.utils_infer import (
|
13 |
+
load_vocoder,
|
14 |
+
load_model,
|
15 |
+
preprocess_ref_audio_text,
|
16 |
+
infer_process,
|
17 |
+
remove_silence_for_generated_wav,
|
18 |
+
)
|
19 |
+
|
20 |
|
21 |
parser = argparse.ArgumentParser(
|
22 |
prog="python3 inference-cli.py",
|
|
|
103 |
spectrogram_path = Path(output_dir)/"out.png"
|
104 |
vocos_local_path = "../checkpoints/charactr/vocos-mel-24khz"
|
105 |
|
106 |
+
vocos = load_vocoder(is_local=args.load_vocoder_from_local, local_path=vocos_local_path)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
109 |
# load models
|
|
|
|
|
|
|
|
|
|
|
110 |
if model == "F5-TTS":
|
111 |
+
model_cls = DiT
|
112 |
+
model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
|
113 |
if ckpt_file == "":
|
114 |
+
repo_name= "F5-TTS"
|
115 |
+
exp_name = "F5TTS_Base"
|
116 |
+
ckpt_step= 1200000
|
117 |
+
ckpt_file = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
|
118 |
+
# ckpt_path = f"ckpts/{exp_name}/model_{ckpt_step}.pt" # .pt | .safetensors; local path
|
|
|
119 |
|
120 |
elif model == "E2-TTS":
|
121 |
+
model_cls = UNetT
|
122 |
+
model_cfg = dict(dim=1024, depth=24, heads=16, ff_mult=4)
|
123 |
if ckpt_file == "":
|
124 |
+
repo_name= "E2-TTS"
|
125 |
+
exp_name = "E2TTS_Base"
|
126 |
+
ckpt_step= 1200000
|
127 |
+
ckpt_file = str(cached_path(f"hf://SWivid/{repo_name}/{exp_name}/model_{ckpt_step}.safetensors"))
|
128 |
+
# ckpt_path = f"ckpts/{exp_name}/model_{ckpt_step}.pt" # .pt | .safetensors; local path
|
129 |
+
|
130 |
+
print(f"Using {model}...")
|
131 |
+
ema_model = load_model(model_cls, model_cfg, ckpt_file, vocab_file)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
|
133 |
|
134 |
+
def main_process(ref_audio, ref_text, text_gen, model_obj, remove_silence):
|
135 |
main_voice = {"ref_audio":ref_audio, "ref_text":ref_text}
|
136 |
if "voices" not in config:
|
137 |
voices = {"main": main_voice}
|
|
|
139 |
voices = config["voices"]
|
140 |
voices["main"] = main_voice
|
141 |
for voice in voices:
|
142 |
+
voices[voice]['ref_audio'], voices[voice]['ref_text'] = preprocess_ref_audio_text(voices[voice]['ref_audio'], voices[voice]['ref_text'])
|
143 |
print("Voice:", voice)
|
144 |
print("Ref_audio:", voices[voice]['ref_audio'])
|
145 |
print("Ref_text:", voices[voice]['ref_text'])
|
|
|
159 |
ref_audio = voices[voice]['ref_audio']
|
160 |
ref_text = voices[voice]['ref_text']
|
161 |
print(f"Voice: {voice}")
|
162 |
+
audio, final_sample_rate, spectragram = infer_process(ref_audio, ref_text, gen_text, model_obj)
|
163 |
generated_audio_segments.append(audio)
|
164 |
|
165 |
if generated_audio_segments:
|
166 |
final_wave = np.concatenate(generated_audio_segments)
|
167 |
with open(wave_path, "wb") as f:
|
168 |
+
sf.write(f.name, final_wave, final_sample_rate)
|
169 |
# Remove silence
|
170 |
if remove_silence:
|
171 |
+
remove_silence_for_generated_wav(f.name)
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
print(f.name)
|
173 |
|
174 |
|
175 |
+
main_process(ref_audio, ref_text, gen_text, ema_model, remove_silence)
|
model/utils_infer.py
ADDED
@@ -0,0 +1,306 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# A unified script for inference process
|
2 |
+
# Make adjustments inside functions, and consider both gradio and cli scripts if need to change func output format
|
3 |
+
|
4 |
+
import re
|
5 |
+
import tempfile
|
6 |
+
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
import torchaudio
|
10 |
+
import tqdm
|
11 |
+
from pydub import AudioSegment, silence
|
12 |
+
from transformers import pipeline
|
13 |
+
from vocos import Vocos
|
14 |
+
|
15 |
+
from model import CFM
|
16 |
+
from model.utils import (
|
17 |
+
load_checkpoint,
|
18 |
+
get_tokenizer,
|
19 |
+
convert_char_to_pinyin,
|
20 |
+
)
|
21 |
+
|
22 |
+
device = (
|
23 |
+
"cuda"
|
24 |
+
if torch.cuda.is_available()
|
25 |
+
else "mps" if torch.backends.mps.is_available() else "cpu"
|
26 |
+
)
|
27 |
+
print(f"Using {device} device")
|
28 |
+
|
29 |
+
asr_pipe = pipeline(
|
30 |
+
"automatic-speech-recognition",
|
31 |
+
model="openai/whisper-large-v3-turbo",
|
32 |
+
torch_dtype=torch.float16,
|
33 |
+
device=device,
|
34 |
+
)
|
35 |
+
|
36 |
+
vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
|
37 |
+
|
38 |
+
|
39 |
+
# -----------------------------------------
|
40 |
+
|
41 |
+
target_sample_rate = 24000
|
42 |
+
n_mel_channels = 100
|
43 |
+
hop_length = 256
|
44 |
+
target_rms = 0.1
|
45 |
+
nfe_step = 32 # 16, 32
|
46 |
+
cfg_strength = 2.0
|
47 |
+
ode_method = "euler"
|
48 |
+
sway_sampling_coef = -1.0
|
49 |
+
speed = 1.0
|
50 |
+
fix_duration = None
|
51 |
+
|
52 |
+
# -----------------------------------------
|
53 |
+
|
54 |
+
|
55 |
+
# chunk text into smaller pieces
|
56 |
+
|
57 |
+
def chunk_text(text, max_chars=135):
|
58 |
+
"""
|
59 |
+
Splits the input text into chunks, each with a maximum number of characters.
|
60 |
+
|
61 |
+
Args:
|
62 |
+
text (str): The text to be split.
|
63 |
+
max_chars (int): The maximum number of characters per chunk.
|
64 |
+
|
65 |
+
Returns:
|
66 |
+
List[str]: A list of text chunks.
|
67 |
+
"""
|
68 |
+
chunks = []
|
69 |
+
current_chunk = ""
|
70 |
+
# Split the text into sentences based on punctuation followed by whitespace
|
71 |
+
sentences = re.split(r'(?<=[;:,.!?])\s+|(?<=[;:,。!?])', text)
|
72 |
+
|
73 |
+
for sentence in sentences:
|
74 |
+
if len(current_chunk.encode('utf-8')) + len(sentence.encode('utf-8')) <= max_chars:
|
75 |
+
current_chunk += sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
|
76 |
+
else:
|
77 |
+
if current_chunk:
|
78 |
+
chunks.append(current_chunk.strip())
|
79 |
+
current_chunk = sentence + " " if sentence and len(sentence[-1].encode('utf-8')) == 1 else sentence
|
80 |
+
|
81 |
+
if current_chunk:
|
82 |
+
chunks.append(current_chunk.strip())
|
83 |
+
|
84 |
+
return chunks
|
85 |
+
|
86 |
+
|
87 |
+
# load vocoder
|
88 |
+
|
89 |
+
def load_vocoder(is_local=False, local_path=""):
|
90 |
+
if is_local:
|
91 |
+
print(f"Load vocos from local path {local_path}")
|
92 |
+
vocos = Vocos.from_hparams(f"{local_path}/config.yaml")
|
93 |
+
state_dict = torch.load(f"{local_path}/pytorch_model.bin", map_location=device)
|
94 |
+
vocos.load_state_dict(state_dict)
|
95 |
+
vocos.eval()
|
96 |
+
else:
|
97 |
+
print("Download Vocos from huggingface charactr/vocos-mel-24khz")
|
98 |
+
vocos = Vocos.from_pretrained("charactr/vocos-mel-24khz")
|
99 |
+
return vocos
|
100 |
+
|
101 |
+
|
102 |
+
# load model for inference
|
103 |
+
|
104 |
+
def load_model(model_cls, model_cfg, ckpt_path, vocab_file=""):
|
105 |
+
|
106 |
+
if vocab_file == "":
|
107 |
+
vocab_file = "Emilia_ZH_EN"
|
108 |
+
tokenizer = "pinyin"
|
109 |
+
else:
|
110 |
+
tokenizer = "custom"
|
111 |
+
|
112 |
+
print("\nvocab : ", vocab_file, tokenizer)
|
113 |
+
print("tokenizer : ", tokenizer)
|
114 |
+
print("model : ", ckpt_path,"\n")
|
115 |
+
|
116 |
+
vocab_char_map, vocab_size = get_tokenizer(vocab_file, tokenizer)
|
117 |
+
model = CFM(
|
118 |
+
transformer=model_cls(
|
119 |
+
**model_cfg, text_num_embeds=vocab_size, mel_dim=n_mel_channels
|
120 |
+
),
|
121 |
+
mel_spec_kwargs=dict(
|
122 |
+
target_sample_rate=target_sample_rate,
|
123 |
+
n_mel_channels=n_mel_channels,
|
124 |
+
hop_length=hop_length,
|
125 |
+
),
|
126 |
+
odeint_kwargs=dict(
|
127 |
+
method=ode_method,
|
128 |
+
),
|
129 |
+
vocab_char_map=vocab_char_map,
|
130 |
+
).to(device)
|
131 |
+
|
132 |
+
model = load_checkpoint(model, ckpt_path, device, use_ema = True)
|
133 |
+
|
134 |
+
return model
|
135 |
+
|
136 |
+
|
137 |
+
# preprocess reference audio and text
|
138 |
+
|
139 |
+
def preprocess_ref_audio_text(ref_audio_orig, ref_text, show_info=print):
|
140 |
+
show_info("Converting audio...")
|
141 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as f:
|
142 |
+
aseg = AudioSegment.from_file(ref_audio_orig)
|
143 |
+
|
144 |
+
non_silent_segs = silence.split_on_silence(
|
145 |
+
aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=1000
|
146 |
+
)
|
147 |
+
non_silent_wave = AudioSegment.silent(duration=0)
|
148 |
+
for non_silent_seg in non_silent_segs:
|
149 |
+
non_silent_wave += non_silent_seg
|
150 |
+
aseg = non_silent_wave
|
151 |
+
|
152 |
+
audio_duration = len(aseg)
|
153 |
+
if audio_duration > 15000:
|
154 |
+
show_info("Audio is over 15s, clipping to only first 15s.")
|
155 |
+
aseg = aseg[:15000]
|
156 |
+
aseg.export(f.name, format="wav")
|
157 |
+
ref_audio = f.name
|
158 |
+
|
159 |
+
if not ref_text.strip():
|
160 |
+
show_info("No reference text provided, transcribing reference audio...")
|
161 |
+
ref_text = asr_pipe(
|
162 |
+
ref_audio,
|
163 |
+
chunk_length_s=30,
|
164 |
+
batch_size=128,
|
165 |
+
generate_kwargs={"task": "transcribe"},
|
166 |
+
return_timestamps=False,
|
167 |
+
)["text"].strip()
|
168 |
+
show_info("Finished transcription")
|
169 |
+
else:
|
170 |
+
show_info("Using custom reference text...")
|
171 |
+
|
172 |
+
# Add the functionality to ensure it ends with ". "
|
173 |
+
if not ref_text.endswith(". ") and not ref_text.endswith("。"):
|
174 |
+
if ref_text.endswith("."):
|
175 |
+
ref_text += " "
|
176 |
+
else:
|
177 |
+
ref_text += ". "
|
178 |
+
|
179 |
+
return ref_audio, ref_text
|
180 |
+
|
181 |
+
|
182 |
+
# infer process: chunk text -> infer batches [i.e. infer_batch_process()]
|
183 |
+
|
184 |
+
def infer_process(ref_audio, ref_text, gen_text, model_obj, cross_fade_duration=0.15, speed=speed, show_info=print, progress=tqdm):
|
185 |
+
|
186 |
+
# Split the input text into batches
|
187 |
+
audio, sr = torchaudio.load(ref_audio)
|
188 |
+
max_chars = int(len(ref_text.encode('utf-8')) / (audio.shape[-1] / sr) * (25 - audio.shape[-1] / sr))
|
189 |
+
gen_text_batches = chunk_text(gen_text, max_chars=max_chars)
|
190 |
+
for i, gen_text in enumerate(gen_text_batches):
|
191 |
+
print(f'gen_text {i}', gen_text)
|
192 |
+
|
193 |
+
show_info(f"Generating audio in {len(gen_text_batches)} batches...")
|
194 |
+
return infer_batch_process((audio, sr), ref_text, gen_text_batches, model_obj, cross_fade_duration, speed, progress)
|
195 |
+
|
196 |
+
|
197 |
+
# infer batches
|
198 |
+
|
199 |
+
def infer_batch_process(ref_audio, ref_text, gen_text_batches, model_obj, cross_fade_duration=0.15, speed=1, progress=tqdm):
|
200 |
+
audio, sr = ref_audio
|
201 |
+
if audio.shape[0] > 1:
|
202 |
+
audio = torch.mean(audio, dim=0, keepdim=True)
|
203 |
+
|
204 |
+
rms = torch.sqrt(torch.mean(torch.square(audio)))
|
205 |
+
if rms < target_rms:
|
206 |
+
audio = audio * target_rms / rms
|
207 |
+
if sr != target_sample_rate:
|
208 |
+
resampler = torchaudio.transforms.Resample(sr, target_sample_rate)
|
209 |
+
audio = resampler(audio)
|
210 |
+
audio = audio.to(device)
|
211 |
+
|
212 |
+
generated_waves = []
|
213 |
+
spectrograms = []
|
214 |
+
|
215 |
+
if len(ref_text[-1].encode('utf-8')) == 1:
|
216 |
+
ref_text = ref_text + " "
|
217 |
+
for i, gen_text in enumerate(progress.tqdm(gen_text_batches)):
|
218 |
+
# Prepare the text
|
219 |
+
text_list = [ref_text + gen_text]
|
220 |
+
final_text_list = convert_char_to_pinyin(text_list)
|
221 |
+
|
222 |
+
# Calculate duration
|
223 |
+
ref_audio_len = audio.shape[-1] // hop_length
|
224 |
+
ref_text_len = len(ref_text.encode('utf-8'))
|
225 |
+
gen_text_len = len(gen_text.encode('utf-8'))
|
226 |
+
duration = ref_audio_len + int(ref_audio_len / ref_text_len * gen_text_len / speed)
|
227 |
+
|
228 |
+
# inference
|
229 |
+
with torch.inference_mode():
|
230 |
+
generated, _ = model_obj.sample(
|
231 |
+
cond=audio,
|
232 |
+
text=final_text_list,
|
233 |
+
duration=duration,
|
234 |
+
steps=nfe_step,
|
235 |
+
cfg_strength=cfg_strength,
|
236 |
+
sway_sampling_coef=sway_sampling_coef,
|
237 |
+
)
|
238 |
+
|
239 |
+
generated = generated.to(torch.float32)
|
240 |
+
generated = generated[:, ref_audio_len:, :]
|
241 |
+
generated_mel_spec = generated.permute(0, 2, 1)
|
242 |
+
generated_wave = vocos.decode(generated_mel_spec.cpu())
|
243 |
+
if rms < target_rms:
|
244 |
+
generated_wave = generated_wave * rms / target_rms
|
245 |
+
|
246 |
+
# wav -> numpy
|
247 |
+
generated_wave = generated_wave.squeeze().cpu().numpy()
|
248 |
+
|
249 |
+
generated_waves.append(generated_wave)
|
250 |
+
spectrograms.append(generated_mel_spec[0].cpu().numpy())
|
251 |
+
|
252 |
+
# Combine all generated waves with cross-fading
|
253 |
+
if cross_fade_duration <= 0:
|
254 |
+
# Simply concatenate
|
255 |
+
final_wave = np.concatenate(generated_waves)
|
256 |
+
else:
|
257 |
+
final_wave = generated_waves[0]
|
258 |
+
for i in range(1, len(generated_waves)):
|
259 |
+
prev_wave = final_wave
|
260 |
+
next_wave = generated_waves[i]
|
261 |
+
|
262 |
+
# Calculate cross-fade samples, ensuring it does not exceed wave lengths
|
263 |
+
cross_fade_samples = int(cross_fade_duration * target_sample_rate)
|
264 |
+
cross_fade_samples = min(cross_fade_samples, len(prev_wave), len(next_wave))
|
265 |
+
|
266 |
+
if cross_fade_samples <= 0:
|
267 |
+
# No overlap possible, concatenate
|
268 |
+
final_wave = np.concatenate([prev_wave, next_wave])
|
269 |
+
continue
|
270 |
+
|
271 |
+
# Overlapping parts
|
272 |
+
prev_overlap = prev_wave[-cross_fade_samples:]
|
273 |
+
next_overlap = next_wave[:cross_fade_samples]
|
274 |
+
|
275 |
+
# Fade out and fade in
|
276 |
+
fade_out = np.linspace(1, 0, cross_fade_samples)
|
277 |
+
fade_in = np.linspace(0, 1, cross_fade_samples)
|
278 |
+
|
279 |
+
# Cross-faded overlap
|
280 |
+
cross_faded_overlap = prev_overlap * fade_out + next_overlap * fade_in
|
281 |
+
|
282 |
+
# Combine
|
283 |
+
new_wave = np.concatenate([
|
284 |
+
prev_wave[:-cross_fade_samples],
|
285 |
+
cross_faded_overlap,
|
286 |
+
next_wave[cross_fade_samples:]
|
287 |
+
])
|
288 |
+
|
289 |
+
final_wave = new_wave
|
290 |
+
|
291 |
+
# Create a combined spectrogram
|
292 |
+
combined_spectrogram = np.concatenate(spectrograms, axis=1)
|
293 |
+
|
294 |
+
return final_wave, target_sample_rate, combined_spectrogram
|
295 |
+
|
296 |
+
|
297 |
+
# remove silence from generated wav
|
298 |
+
|
299 |
+
def remove_silence_for_generated_wav(filename):
|
300 |
+
aseg = AudioSegment.from_file(filename)
|
301 |
+
non_silent_segs = silence.split_on_silence(aseg, min_silence_len=1000, silence_thresh=-50, keep_silence=500)
|
302 |
+
non_silent_wave = AudioSegment.silent(duration=0)
|
303 |
+
for non_silent_seg in non_silent_segs:
|
304 |
+
non_silent_wave += non_silent_seg
|
305 |
+
aseg = non_silent_wave
|
306 |
+
aseg.export(filename, format="wav")
|