Update app.py
Browse files
app.py
CHANGED
|
@@ -86,33 +86,14 @@ os.makedirs("ckpts/Vevo", exist_ok=True)
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
-
|
| 90 |
-
def save_audio_final(waveform, output_path, sample_rate=24000, target_length=None):
|
| 91 |
try:
|
| 92 |
if isinstance(waveform, torch.Tensor):
|
| 93 |
waveform = waveform.detach().cpu()
|
| 94 |
if waveform.dim() == 2 and waveform.shape[0] == 1:
|
| 95 |
waveform = waveform.squeeze(0)
|
| 96 |
waveform = waveform.numpy()
|
| 97 |
-
|
| 98 |
-
# 1. همگامسازی طول (حذف نویز اضافه آخر فایل)
|
| 99 |
-
if target_length is not None:
|
| 100 |
-
if len(waveform) > target_length:
|
| 101 |
-
waveform = waveform[:target_length]
|
| 102 |
-
elif len(waveform) < target_length:
|
| 103 |
-
# اگر کوتاهتر بود، با سکوت پر کن (معمولاً پیش نمیاد)
|
| 104 |
-
padding = np.zeros(target_length - len(waveform))
|
| 105 |
-
waveform = np.concatenate([waveform, padding])
|
| 106 |
-
|
| 107 |
-
# 2. اعمال Fade Out (جلوگیری از صدای کلیک در لحظه قطع شدن)
|
| 108 |
-
fade_len = int(sample_rate * 0.05) # 50 میلی ثانیه
|
| 109 |
-
if len(waveform) > fade_len:
|
| 110 |
-
fade_curve = np.linspace(1, 0, fade_len)
|
| 111 |
-
waveform[-fade_len:] *= fade_curve
|
| 112 |
-
|
| 113 |
-
# 3. ذخیره با فرمت 16 بیتی
|
| 114 |
sf.write(output_path, waveform, sample_rate, subtype='PCM_16')
|
| 115 |
-
|
| 116 |
except Exception as e:
|
| 117 |
print(f"Save error: {e}")
|
| 118 |
raise e
|
|
@@ -142,10 +123,7 @@ inference_pipelines = {}
|
|
| 142 |
def preload_all_resources():
|
| 143 |
print("Preloading resources...")
|
| 144 |
setup_configs()
|
| 145 |
-
|
| 146 |
-
global downloaded_content_style_tokenizer_path
|
| 147 |
-
global downloaded_fmt_path
|
| 148 |
-
global downloaded_vocoder_path
|
| 149 |
|
| 150 |
if not downloaded_resources["tokenizer_vq8192"]:
|
| 151 |
local_dir = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"])
|
|
@@ -196,7 +174,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 196 |
raise ValueError("Please upload audio files")
|
| 197 |
|
| 198 |
try:
|
| 199 |
-
# ---
|
| 200 |
if isinstance(content_wav, tuple):
|
| 201 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 202 |
else:
|
|
@@ -210,12 +188,10 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 210 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 211 |
content_sr = 24000
|
| 212 |
|
|
|
|
| 213 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 214 |
-
|
| 215 |
-
# ذخیره طول دقیق فایل ورودی برای برش نهایی
|
| 216 |
-
target_length_samples = content_tensor.shape[-1]
|
| 217 |
|
| 218 |
-
# ---
|
| 219 |
if isinstance(reference_wav, tuple):
|
| 220 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 221 |
else:
|
|
@@ -231,34 +207,70 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 231 |
|
| 232 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 233 |
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
|
|
|
|
|
|
| 238 |
|
|
|
|
| 239 |
pipeline = get_pipeline()
|
| 240 |
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
)
|
| 246 |
|
| 247 |
-
|
| 248 |
-
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 249 |
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 253 |
|
|
|
|
|
|
|
| 254 |
return output_path
|
| 255 |
|
| 256 |
finally:
|
| 257 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 258 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 259 |
|
| 260 |
-
with gr.Blocks(title="Vevo-Timbre (
|
| 261 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
|
|
|
| 262 |
|
| 263 |
with gr.Row():
|
| 264 |
with gr.Column():
|
|
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
+
def save_audio_pcm16(waveform, output_path, sample_rate=24000):
|
|
|
|
| 90 |
try:
|
| 91 |
if isinstance(waveform, torch.Tensor):
|
| 92 |
waveform = waveform.detach().cpu()
|
| 93 |
if waveform.dim() == 2 and waveform.shape[0] == 1:
|
| 94 |
waveform = waveform.squeeze(0)
|
| 95 |
waveform = waveform.numpy()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
sf.write(output_path, waveform, sample_rate, subtype='PCM_16')
|
|
|
|
| 97 |
except Exception as e:
|
| 98 |
print(f"Save error: {e}")
|
| 99 |
raise e
|
|
|
|
| 123 |
def preload_all_resources():
|
| 124 |
print("Preloading resources...")
|
| 125 |
setup_configs()
|
| 126 |
+
global downloaded_content_style_tokenizer_path, downloaded_fmt_path, downloaded_vocoder_path
|
|
|
|
|
|
|
|
|
|
| 127 |
|
| 128 |
if not downloaded_resources["tokenizer_vq8192"]:
|
| 129 |
local_dir = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"])
|
|
|
|
| 174 |
raise ValueError("Please upload audio files")
|
| 175 |
|
| 176 |
try:
|
| 177 |
+
# --- آماده سازی Content ---
|
| 178 |
if isinstance(content_wav, tuple):
|
| 179 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 180 |
else:
|
|
|
|
| 188 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 189 |
content_sr = 24000
|
| 190 |
|
| 191 |
+
# نرمال سازی
|
| 192 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
|
|
|
|
|
|
|
|
|
| 193 |
|
| 194 |
+
# --- آماده سازی Reference ---
|
| 195 |
if isinstance(reference_wav, tuple):
|
| 196 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 197 |
else:
|
|
|
|
| 207 |
|
| 208 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 209 |
|
| 210 |
+
# اگر رفرنس خیلی طولانی باشد، فقط 20 ثانیه اول کافی است (برای استخراج Timbre)
|
| 211 |
+
# این کار سرعت را بالا میبرد و تاثیری در کیفیت ندارد
|
| 212 |
+
if ref_tensor.shape[1] > 24000 * 20:
|
| 213 |
+
ref_tensor = ref_tensor[:, :24000 * 20]
|
| 214 |
+
|
| 215 |
+
save_audio_pcm16(ref_tensor, temp_reference_path, ref_sr)
|
| 216 |
|
| 217 |
+
# --- منطق Chunking (حل مشکل فایل طولانی) ---
|
| 218 |
pipeline = get_pipeline()
|
| 219 |
|
| 220 |
+
# هر تکه 15 ثانیه (360000 سمپل)
|
| 221 |
+
CHUNK_SIZE = 15 * 24000
|
| 222 |
+
total_samples = content_tensor.shape[1]
|
| 223 |
+
|
| 224 |
+
print(f"[{session_id}] Audio Duration: {total_samples/24000:.2f}s. Starting Chunking...")
|
| 225 |
|
| 226 |
+
generated_chunks = []
|
|
|
|
| 227 |
|
| 228 |
+
for start in range(0, total_samples, CHUNK_SIZE):
|
| 229 |
+
end = min(start + CHUNK_SIZE, total_samples)
|
| 230 |
+
current_chunk = content_tensor[:, start:end]
|
| 231 |
+
|
| 232 |
+
# ذخیره تکه جاری
|
| 233 |
+
save_audio_pcm16(current_chunk, temp_content_path, 24000)
|
| 234 |
+
|
| 235 |
+
print(f"[{session_id}] Processing chunk {start/24000:.1f}s to {end/24000:.1f}s")
|
| 236 |
+
|
| 237 |
+
try:
|
| 238 |
+
gen_chunk = pipeline.inference_fm(
|
| 239 |
+
src_wav_path=temp_content_path,
|
| 240 |
+
timbre_ref_wav_path=temp_reference_path,
|
| 241 |
+
flow_matching_steps=32,
|
| 242 |
+
)
|
| 243 |
+
|
| 244 |
+
# رفع NaN
|
| 245 |
+
if torch.isnan(gen_chunk).any() or torch.isinf(gen_chunk).any():
|
| 246 |
+
gen_chunk = torch.nan_to_num(gen_chunk, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 247 |
+
|
| 248 |
+
# مطمئن شویم تنسور دو بعدی است [1, T]
|
| 249 |
+
if gen_chunk.dim() == 1:
|
| 250 |
+
gen_chunk = gen_chunk.unsqueeze(0)
|
| 251 |
+
|
| 252 |
+
generated_chunks.append(gen_chunk.cpu())
|
| 253 |
+
|
| 254 |
+
except Exception as e:
|
| 255 |
+
print(f"Error in chunk: {e}")
|
| 256 |
+
# اگر خطایی رخ داد، سکوت اضافه کن تا فایل قطع نشود
|
| 257 |
+
silence = torch.zeros((1, end - start))
|
| 258 |
+
generated_chunks.append(silence)
|
| 259 |
+
|
| 260 |
+
# چسباندن تکهها
|
| 261 |
+
final_audio = torch.cat(generated_chunks, dim=1)
|
| 262 |
|
| 263 |
+
# ذخیره نهایی
|
| 264 |
+
save_audio_pcm16(final_audio, output_path, 24000)
|
| 265 |
return output_path
|
| 266 |
|
| 267 |
finally:
|
| 268 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 269 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 270 |
|
| 271 |
+
with gr.Blocks(title="Vevo-Timbre (Long Audio)") as demo:
|
| 272 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 273 |
+
gr.Markdown("پشتیبانی کامل از فایلهای طولانی (بدون نویز و قطعی)")
|
| 274 |
|
| 275 |
with gr.Row():
|
| 276 |
with gr.Column():
|