Spaces:
Running
on
Zero
Running
on
Zero
File size: 14,124 Bytes
6ad1164 384a28c 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 f4f4d73 6ad1164 923bdc3 384a28c 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 6ad1164 923bdc3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 |
#!/usr/bin/env python3
# app.py — ESpeech-TTS с поддержкой ZeroGPU (Hugging Face Spaces)
# ----------------- ZeroGPU / spaces импорт + fallback -----------------
# В среде ZeroGPU доступен пакет `spaces`, который предоставляет декоратор GPU.
# Для локальной отладки мы делаем fallback — noop-декоратор.
import spaces # provided by Spaces/ZeroGPU environment
GPU_DECORATOR = spaces.GPU
print("spaces module available — ZeroGPU features enabled")
import os
import gc
import json
import tempfile
import traceback
from pathlib import Path
import gradio as gr
import numpy as np
import soundfile as sf
import torch
import torchaudio
from huggingface_hub import hf_hub_download
# Ваши зависимости / локальные импорты
from ruaccent import RUAccent
import onnx_asr
from f5_tts.infer.utils_infer import (
infer_process,
load_model,
load_vocoder,
preprocess_ref_audio_text,
remove_silence_for_generated_wav,
save_spectrogram,
tempfile_kwargs,
)
from f5_tts.model import DiT
# Явно включаем ленивый режим кеширования примеров, чтобы примеры не запускались на старте
# (ZeroGPU по умолчанию использует lazy — делаем это явным).
os.environ.setdefault("GRADIO_CACHE_MODE", "lazy")
os.environ.setdefault("GRADIO_CACHE_EXAMPLES", "lazy")
# ----------------- HF hub / модели -----------------
# Настройте репозитории и имена файлов в Hub под себя
MODEL_REPOS = {
"ESpeech-TTS-1 [RL] V2": {
"repo_id": "ESpeech/ESpeech-TTS-1_RL-V2",
"filename": "espeech_tts_rlv2.pt",
},
"ESpeech-TTS-1 [RL] V1": {
"repo_id": "ESpeech/ESpeech-TTS-1_RL-V1",
"filename": "espeech_tts_rlv1.pt",
},
"ESpeech-TTS-1 [SFT] 95K": {
"repo_id": "ESpeech/ESpeech-TTS-1_SFT-95K",
"filename": "espeech_tts_95k.pt",
},
"ESpeech-TTS-1 [SFT] 265K": {
"repo_id": "ESpeech/ESpeech-TTS-1_SFT-256K",
"filename": "espeech_tts_256k.pt",
},
"ESpeech-TTS-1 PODCASTER [SFT]": {
"repo_id": "ESpeech/ESpeech-TTS-1_podcaster",
"filename": "espeech_tts_podcaster.pt",
},
}
# где лежит общий vocab в Hub
VOCAB_REPO = "ESpeech/ESpeech-TTS-1_podcaster"
VOCAB_FILENAME = "vocab.txt"
# токен, если репозитории приватные (в Spaces обычно берут из Secrets)
HF_TOKEN = os.environ.get("HUGGINGFACE_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN") or None
MODEL_CFG = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
# кэш локальных путей после hf_hub_download
_cached_local_paths = {}
loaded_models = {} # хранит объекты моделей в памяти (по имени выбора)
# ----------------- Вспомогательные функции HF -----------------
def hf_download_file(repo_id: str, filename: str, token: str = None):
try:
print(f"hf_hub_download: {repo_id}/{filename}")
p = hf_hub_download(repo_id=repo_id, filename=filename, token=token, repo_type="model")
print(" ->", p)
return p
except Exception as e:
print("Download error:", e)
raise
def get_vocab_path():
key = f"{VOCAB_REPO}::{VOCAB_FILENAME}"
if key in _cached_local_paths and Path(_cached_local_paths[key]).exists():
return _cached_local_paths[key]
p = hf_download_file(VOCAB_REPO, VOCAB_FILENAME, token=HF_TOKEN)
_cached_local_paths[key] = p
return p
def get_model_local_path(choice: str):
if choice not in MODEL_REPOS:
raise KeyError("Unknown model choice: " + repr(choice))
repo = MODEL_REPOS[choice]
key = f"{repo['repo_id']}::{repo['filename']}"
if key in _cached_local_paths and Path(_cached_local_paths[key]).exists():
return _cached_local_paths[key]
p = hf_download_file(repo["repo_id"], repo["filename"], token=HF_TOKEN)
_cached_local_paths[key] = p
return p
def load_model_if_needed(choice: str):
"""
Лениво: если модель уже загружена в loaded_models — вернуть.
Иначе скачать файл (если нужно) и вызвать вашу load_model (возвращает PyTorch модель в CPU).
Не переводим на GPU здесь — это делается внутри GPU-декорированной функции.
"""
if choice in loaded_models:
return loaded_models[choice]
model_file = get_model_local_path(choice)
vocab_file = get_vocab_path()
print(f"Loading model into CPU memory: {choice} from {model_file}")
model = load_model(DiT, MODEL_CFG, model_file, vocab_file=vocab_file)
loaded_models[choice] = model
return model
# ----------------- общие ресурсы (vocoder, RUAccent, ASR) -----------------
print("Loading RUAccent...")
accentizer = RUAccent()
accentizer.load(omograph_model_size='turbo3.1', use_dictionary=True, tiny_mode=False)
print("RUAccent loaded.")
print("Loading ASR (onnx) ...")
asr_model = onnx_asr.load_model("nemo-fastconformer-ru-rnnt")
print("ASR ready.")
print("Loading vocoder (CPU) ...")
vocoder = load_vocoder()
print("Vocoder loaded.")
# ----------------- Основная функция синтеза (GPU-aware) -----------------
# Декорируем synthesize, чтобы при вызове Space выделял GPU (если доступно).
# duration — сколько секунд просим GPU (адаптируйте под ваш инференс).
@GPU_DECORATOR(duration=90)
def synthesize(
model_choice,
ref_audio,
ref_text,
gen_text,
remove_silence,
seed,
cross_fade_duration=0.15,
nfe_step=32,
speed=1.0,
):
"""
Эта функция будет выполняться с выделенным GPU в ZeroGPU Spaces.
Подход:
- лениво загружаем модель (в CPU) если надо
- переносим модель и (если требуется) vocoder на cuda
- делаем infer
- возвращаем модели на CPU и очищаем cuda cache
"""
if not ref_audio:
gr.Warning("Please provide reference audio.")
return None, None, ref_text
if seed is None or seed < 0 or seed > 2**31 - 1:
seed = np.random.randint(0, 2**31 - 1)
torch.manual_seed(int(seed))
if not gen_text or not gen_text.strip():
gr.Warning("Please enter text to generate.")
return None, None, ref_text
# ASR если нужно
if not ref_text or not ref_text.strip():
gr.Info("Reference text is empty. Running ASR to transcribe reference audio...")
try:
waveform, sample_rate = torchaudio.load(ref_audio)
waveform = waveform.numpy()
if waveform.dtype == np.int16:
waveform = waveform / 2**15
elif waveform.dtype == np.int32:
waveform = waveform / 2**31
if waveform.ndim == 2:
waveform = waveform.mean(axis=0)
transcribed_text = asr_model.recognize(waveform, sample_rate=sample_rate)
ref_text = transcribed_text
gr.Info(f"ASR transcription: {ref_text}")
except Exception as e:
gr.Warning(f"ASR failed: {e}")
return None, None, ref_text
# Акцентирование
processed_ref_text = accentizer.process_all(ref_text) if ref_text and ref_text.strip() else ref_text
processed_gen_text = accentizer.process_all(gen_text)
# Ленивая загрузка модели (в CPU)
try:
model = load_model_if_needed(model_choice)
except Exception as e:
gr.Warning(f"Failed to download/load model {model_choice}: {e}")
return None, None, ref_text
# Определяем устройство (в ZeroGPU внутри декоратора должен быть доступен CUDA)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
moved_to_cuda = []
try:
# Переносим модель на GPU (если есть)
if device.type == "cuda":
try:
model.to(device)
moved_to_cuda.append(("model", model))
# если vocoder использует torch — переносим его тоже
try:
vocoder.to(device)
moved_to_cuda.append(("vocoder", vocoder))
except Exception:
# если vocoder не torch-объект — ок
pass
except Exception as e:
print("Warning: failed to move model/vocoder to cuda:", e)
# Препроцессинг рефа (оно ожидает путь/файл)
try:
ref_audio_proc, processed_ref_text = preprocess_ref_audio_text(
ref_audio,
processed_ref_text,
show_info=gr.Info
)
except Exception as e:
gr.Warning(f"Preprocess failed: {e}")
traceback.print_exc()
return None, None, ref_text
# Инференс (предполагается, что infer_process корректно работает и на GPU)
try:
final_wave, final_sample_rate, combined_spectrogram = infer_process(
ref_audio_proc,
processed_ref_text,
processed_gen_text,
model,
vocoder,
cross_fade_duration=cross_fade_duration,
nfe_step=nfe_step,
speed=speed,
show_info=gr.Info,
progress=gr.Progress(),
)
except Exception as e:
gr.Warning(f"Infer failed: {e}")
traceback.print_exc()
return None, None, ref_text
# Удаление тишин (на CPU)
if remove_silence:
try:
with tempfile.NamedTemporaryFile(suffix=".wav", **tempfile_kwargs) as f:
temp_path = f.name
sf.write(temp_path, final_wave, final_sample_rate)
remove_silence_for_generated_wav(temp_path)
final_wave_tensor, _ = torchaudio.load(temp_path)
final_wave = final_wave_tensor.squeeze().cpu().numpy()
except Exception as e:
print("Remove silence failed:", e)
# Сохраняем спектрограмму
try:
with tempfile.NamedTemporaryFile(suffix=".png", **tempfile_kwargs) as tmp_spectrogram:
spectrogram_path = tmp_spectrogram.name
save_spectrogram(combined_spectrogram, spectrogram_path)
except Exception as e:
print("Save spectrogram failed:", e)
spectrogram_path = None
return (final_sample_rate, final_wave), spectrogram_path, processed_ref_text
finally:
# Переносим всё обратно на CPU и очищаем GPU память
if device.type == "cuda":
try:
for name, obj in moved_to_cuda:
try:
obj.to("cpu")
except Exception:
pass
torch.cuda.empty_cache()
# немножко сборки мусора
gc.collect()
except Exception as e:
print("Warning during cuda cleanup:", e)
# ----------------- Gradio UI (как у вас) -----------------
with gr.Blocks(title="ESpeech-TTS (ZeroGPU-ready)") as app:
gr.Markdown("# ESpeech-TTS")
gr.Markdown("Text-to-Speech synthesis system with multiple model variants (models auto-download from HF Hub).")
gr.Markdown("💡 Tip: Leave Reference Text empty to transcribe with ASR. On ZeroGPU the heavy work runs on GPU only during synthesize call.")
model_choice = gr.Dropdown(
choices=list(MODEL_REPOS.keys()),
label="Select Model",
value=list(MODEL_REPOS.keys())[0],
interactive=True
)
with gr.Row():
with gr.Column():
ref_audio_input = gr.Audio(label="Reference Audio", type="filepath")
ref_text_input = gr.Textbox(label="Reference Text", lines=2, placeholder="leave empty → ASR")
with gr.Column():
gen_text_input = gr.Textbox(label="Text to Generate", lines=5, max_lines=20)
with gr.Row():
with gr.Column():
with gr.Accordion("Advanced Settings", open=False):
seed_input = gr.Number(label="Seed (-1 for random)", value=-1, precision=0)
remove_silence = gr.Checkbox(label="Remove Silences", value=False)
speed_slider = gr.Slider(label="Speed", minimum=0.3, maximum=2.0, value=1.0, step=0.1)
nfe_slider = gr.Slider(label="NFE Steps", minimum=4, maximum=64, value=48, step=2)
cross_fade_slider = gr.Slider(label="Cross-Fade Duration (s)", minimum=0.0, maximum=1.0, value=0.15, step=0.01)
generate_btn = gr.Button("🎤 Generate Speech", variant="primary", size="lg")
with gr.Row():
audio_output = gr.Audio(label="Generated Audio", type="numpy")
spectrogram_output = gr.Image(label="Spectrogram", type="filepath")
generate_btn.click(
synthesize,
inputs=[
model_choice,
ref_audio_input,
ref_text_input,
gen_text_input,
remove_silence,
seed_input,
cross_fade_slider,
nfe_slider,
speed_slider,
],
outputs=[audio_output, spectrogram_output, ref_text_input]
)
if __name__ == "__main__":
#app.launch(server_name="0.0.0.0", server_port=7860)
app.launch() |