Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -84,6 +84,10 @@ MODEL_CFG = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_lay
|
|
| 84 |
_cached_local_paths = {}
|
| 85 |
loaded_models = {} # хранит объекты моделей в памяти (по имени выбора)
|
| 86 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 87 |
# ----------------- Вспомогательные функции HF -----------------
|
| 88 |
def hf_download_file(repo_id: str, filename: str, token: str = None):
|
| 89 |
try:
|
|
@@ -143,6 +147,22 @@ print("Loading vocoder (CPU) ...")
|
|
| 143 |
vocoder = load_vocoder()
|
| 144 |
print("Vocoder loaded.")
|
| 145 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
# ----------------- Основная функция синтеза (GPU-aware) -----------------
|
| 147 |
# Декорируем synthesize, чтобы при вызове Space выделял GPU (если доступно).
|
| 148 |
# duration — сколько секунд просим GPU (адаптируйте под ваш инференс).
|
|
@@ -168,7 +188,7 @@ def synthesize(
|
|
| 168 |
"""
|
| 169 |
if not ref_audio:
|
| 170 |
gr.Warning("Please provide reference audio.")
|
| 171 |
-
return None, None, ref_text
|
| 172 |
|
| 173 |
if seed is None or seed < 0 or seed > 2**31 - 1:
|
| 174 |
seed = np.random.randint(0, 2**31 - 1)
|
|
@@ -176,7 +196,7 @@ def synthesize(
|
|
| 176 |
|
| 177 |
if not gen_text or not gen_text.strip():
|
| 178 |
gr.Warning("Please enter text to generate.")
|
| 179 |
-
return None, None, ref_text
|
| 180 |
|
| 181 |
# ASR если нужно
|
| 182 |
if not ref_text or not ref_text.strip():
|
|
@@ -195,18 +215,18 @@ def synthesize(
|
|
| 195 |
gr.Info(f"ASR transcription: {ref_text}")
|
| 196 |
except Exception as e:
|
| 197 |
gr.Warning(f"ASR failed: {e}")
|
| 198 |
-
return None, None, ref_text
|
| 199 |
|
| 200 |
-
# Акцентирование
|
| 201 |
-
processed_ref_text =
|
| 202 |
-
processed_gen_text =
|
| 203 |
|
| 204 |
# Ленивая загрузка модели (в CPU)
|
| 205 |
try:
|
| 206 |
model = load_model_if_needed(model_choice)
|
| 207 |
except Exception as e:
|
| 208 |
gr.Warning(f"Failed to download/load model {model_choice}: {e}")
|
| 209 |
-
return None, None,
|
| 210 |
|
| 211 |
# Определяем устройство (в ZeroGPU внутри декоратора должен быть доступен CUDA)
|
| 212 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
@@ -230,7 +250,7 @@ def synthesize(
|
|
| 230 |
|
| 231 |
# Препроцессинг рефа (оно ожидает путь/файл)
|
| 232 |
try:
|
| 233 |
-
ref_audio_proc,
|
| 234 |
ref_audio,
|
| 235 |
processed_ref_text,
|
| 236 |
show_info=gr.Info
|
|
@@ -238,13 +258,13 @@ def synthesize(
|
|
| 238 |
except Exception as e:
|
| 239 |
gr.Warning(f"Preprocess failed: {e}")
|
| 240 |
traceback.print_exc()
|
| 241 |
-
return None, None,
|
| 242 |
|
| 243 |
# Инференс (предполагается, что infer_process корректно работает и на GPU)
|
| 244 |
try:
|
| 245 |
final_wave, final_sample_rate, combined_spectrogram = infer_process(
|
| 246 |
ref_audio_proc,
|
| 247 |
-
|
| 248 |
processed_gen_text,
|
| 249 |
model,
|
| 250 |
vocoder,
|
|
@@ -257,7 +277,7 @@ def synthesize(
|
|
| 257 |
except Exception as e:
|
| 258 |
gr.Warning(f"Infer failed: {e}")
|
| 259 |
traceback.print_exc()
|
| 260 |
-
return None, None,
|
| 261 |
|
| 262 |
# Удаление тишин (на CPU)
|
| 263 |
if remove_silence:
|
|
@@ -280,7 +300,7 @@ def synthesize(
|
|
| 280 |
print("Save spectrogram failed:", e)
|
| 281 |
spectrogram_path = None
|
| 282 |
|
| 283 |
-
return (final_sample_rate, final_wave), spectrogram_path,
|
| 284 |
|
| 285 |
finally:
|
| 286 |
# Переносим всё обратно на CPU и очищаем GPU память
|
|
@@ -301,6 +321,18 @@ def synthesize(
|
|
| 301 |
with gr.Blocks(title="ESpeech-TTS") as app:
|
| 302 |
gr.Markdown("# ESpeech-TTS")
|
| 303 |
gr.Markdown("See more on https://huggingface.co/ESpeech")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 304 |
|
| 305 |
model_choice = gr.Dropdown(
|
| 306 |
choices=list(MODEL_REPOS.keys()),
|
|
@@ -312,9 +344,29 @@ with gr.Blocks(title="ESpeech-TTS") as app:
|
|
| 312 |
with gr.Row():
|
| 313 |
with gr.Column():
|
| 314 |
ref_audio_input = gr.Audio(label="Reference Audio", type="filepath")
|
| 315 |
-
ref_text_input = gr.Textbox(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 316 |
with gr.Column():
|
| 317 |
-
gen_text_input = gr.Textbox(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 318 |
|
| 319 |
with gr.Row():
|
| 320 |
with gr.Column():
|
|
@@ -331,6 +383,37 @@ with gr.Blocks(title="ESpeech-TTS") as app:
|
|
| 331 |
audio_output = gr.Audio(label="Generated Audio", type="numpy")
|
| 332 |
spectrogram_output = gr.Image(label="Spectrogram", type="filepath")
|
| 333 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 334 |
generate_btn.click(
|
| 335 |
synthesize,
|
| 336 |
inputs=[
|
|
@@ -344,7 +427,7 @@ with gr.Blocks(title="ESpeech-TTS") as app:
|
|
| 344 |
nfe_slider,
|
| 345 |
speed_slider,
|
| 346 |
],
|
| 347 |
-
outputs=[audio_output, spectrogram_output,
|
| 348 |
)
|
| 349 |
|
| 350 |
if __name__ == "__main__":
|
|
|
|
| 84 |
_cached_local_paths = {}
|
| 85 |
loaded_models = {} # хранит объекты моделей в памяти (по имени выбора)
|
| 86 |
|
| 87 |
+
# Пример текста для демонстрации
|
| 88 |
+
EXAMPLE_TEXT = "Экспериментальный центр напоминает вам о том, что кубы не умеют разговаривать. В случае, если грузовой куб все же заговорит, центр настоятельно рекомендует вам игнорировать его советы."
|
| 89 |
+
EXAMPLE_REF_AUDIO = "ref/example.mp3"
|
| 90 |
+
|
| 91 |
# ----------------- Вспомогательные функции HF -----------------
|
| 92 |
def hf_download_file(repo_id: str, filename: str, token: str = None):
|
| 93 |
try:
|
|
|
|
| 147 |
vocoder = load_vocoder()
|
| 148 |
print("Vocoder loaded.")
|
| 149 |
|
| 150 |
+
# ----------------- Функция для обработки текста с учетом "+" -----------------
|
| 151 |
+
def process_text_with_accent(text, accentizer):
|
| 152 |
+
"""
|
| 153 |
+
Обрабатывает текст через RUAccent, если в нем нет символа '+'.
|
| 154 |
+
Если есть '+' - пользователь сам проставил ударения, не трогаем.
|
| 155 |
+
"""
|
| 156 |
+
if not text or not text.strip():
|
| 157 |
+
return text
|
| 158 |
+
|
| 159 |
+
if '+' in text:
|
| 160 |
+
# Пользователь сам проставил ударения
|
| 161 |
+
return text
|
| 162 |
+
else:
|
| 163 |
+
# Прогоняем через RUAccent
|
| 164 |
+
return accentizer.process_all(text)
|
| 165 |
+
|
| 166 |
# ----------------- Основная функция синтеза (GPU-aware) -----------------
|
| 167 |
# Декорируем synthesize, чтобы при вызове Space выделял GPU (если доступно).
|
| 168 |
# duration — сколько секунд просим GPU (адаптируйте под ваш инференс).
|
|
|
|
| 188 |
"""
|
| 189 |
if not ref_audio:
|
| 190 |
gr.Warning("Please provide reference audio.")
|
| 191 |
+
return None, None, ref_text, gen_text
|
| 192 |
|
| 193 |
if seed is None or seed < 0 or seed > 2**31 - 1:
|
| 194 |
seed = np.random.randint(0, 2**31 - 1)
|
|
|
|
| 196 |
|
| 197 |
if not gen_text or not gen_text.strip():
|
| 198 |
gr.Warning("Please enter text to generate.")
|
| 199 |
+
return None, None, ref_text, gen_text
|
| 200 |
|
| 201 |
# ASR если нужно
|
| 202 |
if not ref_text or not ref_text.strip():
|
|
|
|
| 215 |
gr.Info(f"ASR transcription: {ref_text}")
|
| 216 |
except Exception as e:
|
| 217 |
gr.Warning(f"ASR failed: {e}")
|
| 218 |
+
return None, None, ref_text, gen_text
|
| 219 |
|
| 220 |
+
# Акцентирование с учетом наличия символа "+"
|
| 221 |
+
processed_ref_text = process_text_with_accent(ref_text, accentizer)
|
| 222 |
+
processed_gen_text = process_text_with_accent(gen_text, accentizer)
|
| 223 |
|
| 224 |
# Ленивая загрузка модели (в CPU)
|
| 225 |
try:
|
| 226 |
model = load_model_if_needed(model_choice)
|
| 227 |
except Exception as e:
|
| 228 |
gr.Warning(f"Failed to download/load model {model_choice}: {e}")
|
| 229 |
+
return None, None, processed_ref_text, processed_gen_text
|
| 230 |
|
| 231 |
# Определяем устройство (в ZeroGPU внутри декоратора должен быть доступен CUDA)
|
| 232 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
|
| 250 |
|
| 251 |
# Препроцессинг рефа (оно ожидает путь/файл)
|
| 252 |
try:
|
| 253 |
+
ref_audio_proc, processed_ref_text_final = preprocess_ref_audio_text(
|
| 254 |
ref_audio,
|
| 255 |
processed_ref_text,
|
| 256 |
show_info=gr.Info
|
|
|
|
| 258 |
except Exception as e:
|
| 259 |
gr.Warning(f"Preprocess failed: {e}")
|
| 260 |
traceback.print_exc()
|
| 261 |
+
return None, None, processed_ref_text, processed_gen_text
|
| 262 |
|
| 263 |
# Инференс (предполагается, что infer_process корректно работает и на GPU)
|
| 264 |
try:
|
| 265 |
final_wave, final_sample_rate, combined_spectrogram = infer_process(
|
| 266 |
ref_audio_proc,
|
| 267 |
+
processed_ref_text_final,
|
| 268 |
processed_gen_text,
|
| 269 |
model,
|
| 270 |
vocoder,
|
|
|
|
| 277 |
except Exception as e:
|
| 278 |
gr.Warning(f"Infer failed: {e}")
|
| 279 |
traceback.print_exc()
|
| 280 |
+
return None, None, processed_ref_text, processed_gen_text
|
| 281 |
|
| 282 |
# Удаление тишин (на CPU)
|
| 283 |
if remove_silence:
|
|
|
|
| 300 |
print("Save spectrogram failed:", e)
|
| 301 |
spectrogram_path = None
|
| 302 |
|
| 303 |
+
return (final_sample_rate, final_wave), spectrogram_path, processed_ref_text_final, processed_gen_text
|
| 304 |
|
| 305 |
finally:
|
| 306 |
# Переносим всё обратно на CPU и очищаем GPU память
|
|
|
|
| 321 |
with gr.Blocks(title="ESpeech-TTS") as app:
|
| 322 |
gr.Markdown("# ESpeech-TTS")
|
| 323 |
gr.Markdown("See more on https://huggingface.co/ESpeech")
|
| 324 |
+
gr.Markdown("💡 **Tip:** Add '+' symbol in text to mark custom stress (e.g., 'прив+ет'). Text with '+' won't be processed by RUAccent.")
|
| 325 |
+
|
| 326 |
+
# Описание моделей на русском языке
|
| 327 |
+
gr.Markdown("""
|
| 328 |
+
## 📋 Описание моделей:
|
| 329 |
+
|
| 330 |
+
- **ESpeech-TTS-1 [RL] V1** - Первая версия модели с RL
|
| 331 |
+
- **ESpeech-TTS-1 [RL] V2** - Вторая версия модели с RL
|
| 332 |
+
- **ESpeech-TTS-1 PODCASTER [SFT]** - Модель обученная только на подкастах, лучше генерирует спонтанную речь
|
| 333 |
+
- **ESpeech-TTS-1 [SFT] 95K** - чекпоинт с 95000 шагов (на нем основана RL V1)
|
| 334 |
+
- **ESpeech-TTS-1 [SFT] 265K** - чекпоинт с 265000 шагов (на нем основана RL V2)
|
| 335 |
+
""")
|
| 336 |
|
| 337 |
model_choice = gr.Dropdown(
|
| 338 |
choices=list(MODEL_REPOS.keys()),
|
|
|
|
| 344 |
with gr.Row():
|
| 345 |
with gr.Column():
|
| 346 |
ref_audio_input = gr.Audio(label="Reference Audio", type="filepath")
|
| 347 |
+
ref_text_input = gr.Textbox(
|
| 348 |
+
label="Reference Text",
|
| 349 |
+
lines=2,
|
| 350 |
+
placeholder="leave empty → ASR will transcribe"
|
| 351 |
+
)
|
| 352 |
+
ref_text_output = gr.Textbox(
|
| 353 |
+
label="Processed Reference Text (with accents)",
|
| 354 |
+
lines=2,
|
| 355 |
+
interactive=False
|
| 356 |
+
)
|
| 357 |
with gr.Column():
|
| 358 |
+
gen_text_input = gr.Textbox(
|
| 359 |
+
label="Text to Generate",
|
| 360 |
+
lines=5,
|
| 361 |
+
max_lines=20,
|
| 362 |
+
placeholder="Enter text to synthesize..."
|
| 363 |
+
)
|
| 364 |
+
gen_text_output = gr.Textbox(
|
| 365 |
+
label="Processed Text to Generate (with accents)",
|
| 366 |
+
lines=5,
|
| 367 |
+
max_lines=20,
|
| 368 |
+
interactive=False
|
| 369 |
+
)
|
| 370 |
|
| 371 |
with gr.Row():
|
| 372 |
with gr.Column():
|
|
|
|
| 383 |
audio_output = gr.Audio(label="Generated Audio", type="numpy")
|
| 384 |
spectrogram_output = gr.Image(label="Spectrogram", type="filepath")
|
| 385 |
|
| 386 |
+
# Примеры
|
| 387 |
+
gr.Markdown("## 🎯 Example")
|
| 388 |
+
gr.Examples(
|
| 389 |
+
examples=[
|
| 390 |
+
[
|
| 391 |
+
EXAMPLE_REF_AUDIO, # ref_audio
|
| 392 |
+
"", # ref_text (empty for ASR)
|
| 393 |
+
EXAMPLE_TEXT, # gen_text
|
| 394 |
+
False, # remove_silence
|
| 395 |
+
42, # seed
|
| 396 |
+
0.15, # cross_fade
|
| 397 |
+
48, # nfe_step
|
| 398 |
+
1.0, # speed
|
| 399 |
+
]
|
| 400 |
+
],
|
| 401 |
+
inputs=[
|
| 402 |
+
ref_audio_input,
|
| 403 |
+
ref_text_input,
|
| 404 |
+
gen_text_input,
|
| 405 |
+
remove_silence,
|
| 406 |
+
seed_input,
|
| 407 |
+
cross_fade_slider,
|
| 408 |
+
nfe_slider,
|
| 409 |
+
speed_slider,
|
| 410 |
+
],
|
| 411 |
+
outputs=[audio_output, spectrogram_output, ref_text_output, gen_text_output],
|
| 412 |
+
fn=lambda *args: synthesize(model_choice.value, *args),
|
| 413 |
+
cache_examples=True,
|
| 414 |
+
run_on_click=True,
|
| 415 |
+
)
|
| 416 |
+
|
| 417 |
generate_btn.click(
|
| 418 |
synthesize,
|
| 419 |
inputs=[
|
|
|
|
| 427 |
nfe_slider,
|
| 428 |
speed_slider,
|
| 429 |
],
|
| 430 |
+
outputs=[audio_output, spectrogram_output, ref_text_output, gen_text_output]
|
| 431 |
)
|
| 432 |
|
| 433 |
if __name__ == "__main__":
|