Update app.py
Browse files
app.py
CHANGED
|
@@ -12,7 +12,7 @@ import subprocess
|
|
| 12 |
import re
|
| 13 |
import spaces
|
| 14 |
import uuid
|
| 15 |
-
import soundfile as sf
|
| 16 |
|
| 17 |
# فقط منابع ضروری
|
| 18 |
downloaded_resources = {
|
|
@@ -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):
|
|
@@ -169,7 +169,6 @@ def get_pipeline():
|
|
| 169 |
|
| 170 |
@spaces.GPU()
|
| 171 |
def vevo_timbre(content_wav, reference_wav):
|
| 172 |
-
# 1. ایجاد نام یکتا برای هر کاربر (جلوگیری از قاطی شدن فایلها)
|
| 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"
|
|
@@ -179,7 +178,6 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 179 |
raise ValueError("Please upload audio files")
|
| 180 |
|
| 181 |
try:
|
| 182 |
-
# --- پردازش و نرمالسازی صداها ---
|
| 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:
|
|
@@ -189,15 +187,12 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 189 |
content_data = np.mean(content_data, axis=1)
|
| 190 |
|
| 191 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
| 192 |
-
# مهم: استفاده از torchaudio برای ریسمپل دقیق (جلوگیری از نویز)
|
| 193 |
if content_sr != 24000:
|
| 194 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 195 |
content_sr = 24000
|
| 196 |
|
| 197 |
-
# نرمالسازی صدا (خیلی مهم برای کیفیت)
|
| 198 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 199 |
|
| 200 |
-
# --- پردازش 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:
|
|
@@ -213,15 +208,14 @@ 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 |
-
# استفاده از soundfile برای ذخیره
|
| 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 Audio
|
| 221 |
|
| 222 |
pipeline = get_pipeline()
|
| 223 |
|
| 224 |
-
# اجرای مدل روی کل فایل (بدون تکه تکه کردن)
|
| 225 |
gen_audio = pipeline.inference_fm(
|
| 226 |
src_wav_path=temp_content_path,
|
| 227 |
timbre_ref_wav_path=temp_reference_path,
|
|
@@ -236,7 +230,6 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 236 |
return output_path
|
| 237 |
|
| 238 |
finally:
|
| 239 |
-
# پاکسازی فایلهای موقت
|
| 240 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 241 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 242 |
|
|
|
|
| 12 |
import re
|
| 13 |
import spaces
|
| 14 |
import uuid
|
| 15 |
+
import soundfile as sf
|
| 16 |
|
| 17 |
# فقط منابع ضروری
|
| 18 |
downloaded_resources = {
|
|
|
|
| 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 |
|
| 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 |
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 |
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 |
|
| 209 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 210 |
|
| 211 |
+
# استفاده از soundfile برای ذخیره موقت
|
| 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 Audio...")
|
| 216 |
|
| 217 |
pipeline = get_pipeline()
|
| 218 |
|
|
|
|
| 219 |
gen_audio = pipeline.inference_fm(
|
| 220 |
src_wav_path=temp_content_path,
|
| 221 |
timbre_ref_wav_path=temp_reference_path,
|
|
|
|
| 230 |
return output_path
|
| 231 |
|
| 232 |
finally:
|
|
|
|
| 233 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 234 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 235 |
|