Update app.py
Browse files
app.py
CHANGED
|
@@ -86,9 +86,8 @@ os.makedirs("ckpts/Vevo", exist_ok=True)
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
-
# --- تابع ذخیره سازی
|
| 90 |
-
|
| 91 |
-
def save_audio_pcm16(waveform, output_path, sample_rate=24000):
|
| 92 |
try:
|
| 93 |
if isinstance(waveform, torch.Tensor):
|
| 94 |
waveform = waveform.detach().cpu()
|
|
@@ -96,7 +95,22 @@ def save_audio_pcm16(waveform, output_path, sample_rate=24000):
|
|
| 96 |
waveform = waveform.squeeze(0)
|
| 97 |
waveform = waveform.numpy()
|
| 98 |
|
| 99 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
sf.write(output_path, waveform, sample_rate, subtype='PCM_16')
|
| 101 |
|
| 102 |
except Exception as e:
|
|
@@ -192,12 +206,14 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 192 |
content_data = np.mean(content_data, axis=1)
|
| 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
|
| 199 |
|
| 200 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
|
|
|
|
|
|
|
|
|
| 201 |
|
| 202 |
# --- پردازش صدای رفرنس ---
|
| 203 |
if isinstance(reference_wav, tuple):
|
|
@@ -215,15 +231,13 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 215 |
|
| 216 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 217 |
|
| 218 |
-
|
| 219 |
-
|
| 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,
|
|
@@ -233,15 +247,17 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 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 |
-
#
|
| 237 |
-
|
|
|
|
|
|
|
| 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 (
|
| 245 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 246 |
|
| 247 |
with gr.Row():
|
|
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
+
# --- تابع ذخیره سازی پیشرفته (حذف نویز + فرمت استاندارد) ---
|
| 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()
|
|
|
|
| 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:
|
|
|
|
| 206 |
content_data = np.mean(content_data, axis=1)
|
| 207 |
|
| 208 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
|
|
|
| 209 |
if content_sr != 24000:
|
| 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 |
# --- پردازش صدای رفرنس ---
|
| 219 |
if isinstance(reference_wav, tuple):
|
|
|
|
| 231 |
|
| 232 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 233 |
|
| 234 |
+
save_audio_final(content_tensor, temp_content_path, content_sr)
|
| 235 |
+
save_audio_final(ref_tensor, temp_reference_path, ref_sr)
|
|
|
|
| 236 |
|
| 237 |
print(f"[{session_id}] Processing...")
|
| 238 |
|
| 239 |
pipeline = get_pipeline()
|
| 240 |
|
|
|
|
| 241 |
gen_audio = pipeline.inference_fm(
|
| 242 |
src_wav_path=temp_content_path,
|
| 243 |
timbre_ref_wav_path=temp_reference_path,
|
|
|
|
| 247 |
if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any():
|
| 248 |
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 249 |
|
| 250 |
+
# اینجا فایل را دقیقاً به اندازه ورودی برش میزنیم
|
| 251 |
+
# این کار باعث میشود نویز اضافهای که مدل در پایان تولید کرده حذف شود
|
| 252 |
+
save_audio_final(gen_audio, output_path, 24000, target_length=target_length_samples)
|
| 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 (Clean)") as demo:
|
| 261 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 262 |
|
| 263 |
with gr.Row():
|