Update app.py
Browse files
app.py
CHANGED
|
@@ -14,7 +14,7 @@ import spaces
|
|
| 14 |
import uuid
|
| 15 |
import soundfile as sf
|
| 16 |
|
| 17 |
-
#
|
| 18 |
downloaded_resources = {
|
| 19 |
"configs": False,
|
| 20 |
"tokenizer_vq8192": False,
|
|
@@ -86,7 +86,7 @@ os.makedirs("ckpts/Vevo", exist_ok=True)
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
-
# تابع ذخیره سازی امن
|
| 90 |
def my_save_audio(waveform, output_path, sample_rate=24000):
|
| 91 |
try:
|
| 92 |
if isinstance(waveform, torch.Tensor):
|
|
@@ -169,6 +169,7 @@ def get_pipeline():
|
|
| 169 |
|
| 170 |
@spaces.GPU()
|
| 171 |
def vevo_timbre(content_wav, reference_wav):
|
|
|
|
| 172 |
session_id = str(uuid.uuid4())[:8]
|
| 173 |
temp_content_path = f"wav/c_{session_id}.wav"
|
| 174 |
temp_reference_path = f"wav/r_{session_id}.wav"
|
|
@@ -178,6 +179,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 178 |
raise ValueError("Please upload audio files")
|
| 179 |
|
| 180 |
try:
|
|
|
|
| 181 |
if isinstance(content_wav, tuple):
|
| 182 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 183 |
else:
|
|
@@ -187,12 +189,15 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 187 |
content_data = np.mean(content_data, axis=1)
|
| 188 |
|
| 189 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
|
|
|
|
|
|
| 190 |
if content_sr != 24000:
|
| 191 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 192 |
content_sr = 24000
|
| 193 |
|
| 194 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 195 |
|
|
|
|
| 196 |
if isinstance(reference_wav, tuple):
|
| 197 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 198 |
else:
|
|
@@ -208,11 +213,11 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 208 |
|
| 209 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 210 |
|
| 211 |
-
#
|
| 212 |
sf.write(temp_content_path, content_tensor.squeeze().cpu().numpy(), content_sr)
|
| 213 |
sf.write(temp_reference_path, ref_tensor.squeeze().cpu().numpy(), ref_sr)
|
| 214 |
|
| 215 |
-
print(f"[{session_id}] Processing
|
| 216 |
|
| 217 |
pipeline = get_pipeline()
|
| 218 |
|
|
@@ -226,6 +231,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 226 |
print("Warning: NaN fixed")
|
| 227 |
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 228 |
|
|
|
|
| 229 |
my_save_audio(gen_audio, output_path=output_path)
|
| 230 |
return output_path
|
| 231 |
|
|
|
|
| 14 |
import uuid
|
| 15 |
import soundfile as sf
|
| 16 |
|
| 17 |
+
# منابع ضروری
|
| 18 |
downloaded_resources = {
|
| 19 |
"configs": False,
|
| 20 |
"tokenizer_vq8192": False,
|
|
|
|
| 86 |
|
| 87 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
|
| 89 |
+
# تابع ذخیره سازی امن (جایگزین torchaudio)
|
| 90 |
def my_save_audio(waveform, output_path, sample_rate=24000):
|
| 91 |
try:
|
| 92 |
if isinstance(waveform, torch.Tensor):
|
|
|
|
| 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"
|
|
|
|
| 179 |
raise ValueError("Please upload audio files")
|
| 180 |
|
| 181 |
try:
|
| 182 |
+
# --- پردازش صدای اصلی ---
|
| 183 |
if isinstance(content_wav, tuple):
|
| 184 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 185 |
else:
|
|
|
|
| 189 |
content_data = np.mean(content_data, axis=1)
|
| 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
|
| 197 |
|
| 198 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 199 |
|
| 200 |
+
# --- پردازش صدای رفرنس ---
|
| 201 |
if isinstance(reference_wav, tuple):
|
| 202 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 203 |
else:
|
|
|
|
| 213 |
|
| 214 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 215 |
|
| 216 |
+
# ذخیره موقت با soundfile (برای جلوگیری از ارور TorchCodec)
|
| 217 |
sf.write(temp_content_path, content_tensor.squeeze().cpu().numpy(), content_sr)
|
| 218 |
sf.write(temp_reference_path, ref_tensor.squeeze().cpu().numpy(), ref_sr)
|
| 219 |
|
| 220 |
+
print(f"[{session_id}] Processing...")
|
| 221 |
|
| 222 |
pipeline = get_pipeline()
|
| 223 |
|
|
|
|
| 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 با soundfile
|
| 235 |
my_save_audio(gen_audio, output_path=output_path)
|
| 236 |
return output_path
|
| 237 |
|