Update app.py
Browse files
app.py
CHANGED
|
@@ -86,15 +86,19 @@ os.makedirs("ckpts/Vevo", exist_ok=True)
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
-
# تابع ذخیره سازی
|
| 90 |
-
|
|
|
|
| 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 |
except Exception as e:
|
| 99 |
print(f"Save error: {e}")
|
| 100 |
raise e
|
|
@@ -169,7 +173,6 @@ def get_pipeline():
|
|
| 169 |
|
| 170 |
@spaces.GPU()
|
| 171 |
def vevo_timbre(content_wav, reference_wav):
|
| 172 |
-
# تولید نام فایل امن
|
| 173 |
session_id = str(uuid.uuid4())[:8]
|
| 174 |
temp_content_path = f"wav/c_{session_id}.wav"
|
| 175 |
temp_reference_path = f"wav/r_{session_id}.wav"
|
|
@@ -190,7 +193,6 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 190 |
|
| 191 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
| 192 |
|
| 193 |
-
# ریسمپل با torchaudio (اینجا ارور نمیده چون ذخیره نمیکنیم، فقط پردازش میکنیم)
|
| 194 |
if content_sr != 24000:
|
| 195 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 196 |
content_sr = 24000
|
|
@@ -213,14 +215,15 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 213 |
|
| 214 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 215 |
|
| 216 |
-
# ذخیره
|
| 217 |
-
|
| 218 |
-
|
| 219 |
|
| 220 |
print(f"[{session_id}] Processing...")
|
| 221 |
|
| 222 |
pipeline = get_pipeline()
|
| 223 |
|
|
|
|
| 224 |
gen_audio = pipeline.inference_fm(
|
| 225 |
src_wav_path=temp_content_path,
|
| 226 |
timbre_ref_wav_path=temp_reference_path,
|
|
@@ -228,18 +231,17 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 228 |
)
|
| 229 |
|
| 230 |
if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any():
|
| 231 |
-
print("Warning: NaN fixed")
|
| 232 |
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 233 |
|
| 234 |
-
# ذخیره نهایی
|
| 235 |
-
|
| 236 |
return output_path
|
| 237 |
|
| 238 |
finally:
|
| 239 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 240 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 241 |
|
| 242 |
-
with gr.Blocks(title="Vevo-Timbre (
|
| 243 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 244 |
|
| 245 |
with gr.Row():
|
|
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
+
# --- تابع ذخیره سازی دقیق (16-bit PCM) ---
|
| 90 |
+
# این تابع کلید حل مشکل نویز صداست. فایل را دقیقاً مثل WAV استاندارد ذخیره میکند.
|
| 91 |
+
def save_audio_pcm16(waveform, output_path, sample_rate=24000):
|
| 92 |
try:
|
| 93 |
if isinstance(waveform, torch.Tensor):
|
| 94 |
waveform = waveform.detach().cpu()
|
| 95 |
if waveform.dim() == 2 and waveform.shape[0] == 1:
|
| 96 |
waveform = waveform.squeeze(0)
|
| 97 |
waveform = waveform.numpy()
|
| 98 |
+
|
| 99 |
+
# تبدیل به فرمت 16 بیتی برای جلوگیری از نویز
|
| 100 |
+
sf.write(output_path, waveform, sample_rate, subtype='PCM_16')
|
| 101 |
+
|
| 102 |
except Exception as e:
|
| 103 |
print(f"Save error: {e}")
|
| 104 |
raise e
|
|
|
|
| 173 |
|
| 174 |
@spaces.GPU()
|
| 175 |
def vevo_timbre(content_wav, reference_wav):
|
|
|
|
| 176 |
session_id = str(uuid.uuid4())[:8]
|
| 177 |
temp_content_path = f"wav/c_{session_id}.wav"
|
| 178 |
temp_reference_path = f"wav/r_{session_id}.wav"
|
|
|
|
| 193 |
|
| 194 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
| 195 |
|
|
|
|
| 196 |
if content_sr != 24000:
|
| 197 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 198 |
content_sr = 24000
|
|
|
|
| 215 |
|
| 216 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 217 |
|
| 218 |
+
# *** ذخیره با فرمت PCM_16 (کلید حل مشکل نویز) ***
|
| 219 |
+
save_audio_pcm16(content_tensor, temp_content_path, content_sr)
|
| 220 |
+
save_audio_pcm16(ref_tensor, temp_reference_path, ref_sr)
|
| 221 |
|
| 222 |
print(f"[{session_id}] Processing...")
|
| 223 |
|
| 224 |
pipeline = get_pipeline()
|
| 225 |
|
| 226 |
+
# اجرای مدل
|
| 227 |
gen_audio = pipeline.inference_fm(
|
| 228 |
src_wav_path=temp_content_path,
|
| 229 |
timbre_ref_wav_path=temp_reference_path,
|
|
|
|
| 231 |
)
|
| 232 |
|
| 233 |
if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any():
|
|
|
|
| 234 |
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 235 |
|
| 236 |
+
# ذخیره خروجی نهایی
|
| 237 |
+
save_audio_pcm16(gen_audio, output_path, 24000)
|
| 238 |
return output_path
|
| 239 |
|
| 240 |
finally:
|
| 241 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 242 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 243 |
|
| 244 |
+
with gr.Blocks(title="Vevo-Timbre (High Quality)") as demo:
|
| 245 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 246 |
|
| 247 |
with gr.Row():
|