Update app.py
Browse files
app.py
CHANGED
|
@@ -12,8 +12,9 @@ import subprocess
|
|
| 12 |
import re
|
| 13 |
import spaces
|
| 14 |
import uuid
|
|
|
|
| 15 |
|
| 16 |
-
#
|
| 17 |
downloaded_resources = {
|
| 18 |
"configs": False,
|
| 19 |
"tokenizer_vq8192": False,
|
|
@@ -28,8 +29,6 @@ def install_espeak():
|
|
| 28 |
print("Installing espeak-ng...")
|
| 29 |
subprocess.run(["apt-get", "update"], check=True)
|
| 30 |
subprocess.run(["apt-get", "install", "-y", "espeak-ng", "espeak-ng-data"], check=True)
|
| 31 |
-
else:
|
| 32 |
-
print("espeak-ng is installed.")
|
| 33 |
except Exception as e:
|
| 34 |
print(f"Error installing espeak-ng: {e}")
|
| 35 |
|
|
@@ -85,8 +84,20 @@ if os.path.dirname(os.path.abspath("Amphion")) not in sys.path:
|
|
| 85 |
os.makedirs("wav", exist_ok=True)
|
| 86 |
os.makedirs("ckpts/Vevo", exist_ok=True)
|
| 87 |
|
| 88 |
-
|
| 89 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
|
| 91 |
def setup_configs():
|
| 92 |
if downloaded_resources["configs"]: return
|
|
@@ -158,7 +169,7 @@ def get_pipeline():
|
|
| 158 |
|
| 159 |
@spaces.GPU()
|
| 160 |
def vevo_timbre(content_wav, reference_wav):
|
| 161 |
-
# ایجاد نام
|
| 162 |
session_id = str(uuid.uuid4())[:8]
|
| 163 |
temp_content_path = f"wav/c_{session_id}.wav"
|
| 164 |
temp_reference_path = f"wav/r_{session_id}.wav"
|
|
@@ -168,7 +179,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 168 |
raise ValueError("Please upload audio files")
|
| 169 |
|
| 170 |
try:
|
| 171 |
-
# --- پردازش
|
| 172 |
if isinstance(content_wav, tuple):
|
| 173 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 174 |
else:
|
|
@@ -178,13 +189,15 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 178 |
content_data = np.mean(content_data, axis=1)
|
| 179 |
|
| 180 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
|
|
|
| 181 |
if content_sr != 24000:
|
| 182 |
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 183 |
content_sr = 24000
|
| 184 |
|
|
|
|
| 185 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 186 |
|
| 187 |
-
# --- پردازش
|
| 188 |
if isinstance(reference_wav, tuple):
|
| 189 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 190 |
else:
|
|
@@ -200,15 +213,15 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 200 |
|
| 201 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 202 |
|
| 203 |
-
#
|
| 204 |
-
|
| 205 |
-
|
| 206 |
|
| 207 |
-
print(f"[{session_id}] Processing Audio...")
|
| 208 |
|
| 209 |
pipeline = get_pipeline()
|
| 210 |
|
| 211 |
-
# اجرای مدل روی کل فایل (بدون تکه تکه کردن
|
| 212 |
gen_audio = pipeline.inference_fm(
|
| 213 |
src_wav_path=temp_content_path,
|
| 214 |
timbre_ref_wav_path=temp_reference_path,
|
|
@@ -219,7 +232,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 219 |
print("Warning: NaN fixed")
|
| 220 |
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 221 |
|
| 222 |
-
|
| 223 |
return output_path
|
| 224 |
|
| 225 |
finally:
|
|
@@ -227,7 +240,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 227 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 228 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 229 |
|
| 230 |
-
with gr.Blocks(title="Vevo-Timbre (
|
| 231 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 232 |
|
| 233 |
with gr.Row():
|
|
|
|
| 12 |
import re
|
| 13 |
import spaces
|
| 14 |
import uuid
|
| 15 |
+
import soundfile as sf # استفاده مستقیم برای حل مشکل ذخیرهسازی
|
| 16 |
|
| 17 |
+
# فقط منابع ضروری
|
| 18 |
downloaded_resources = {
|
| 19 |
"configs": False,
|
| 20 |
"tokenizer_vq8192": False,
|
|
|
|
| 29 |
print("Installing espeak-ng...")
|
| 30 |
subprocess.run(["apt-get", "update"], check=True)
|
| 31 |
subprocess.run(["apt-get", "install", "-y", "espeak-ng", "espeak-ng-data"], check=True)
|
|
|
|
|
|
|
| 32 |
except Exception as e:
|
| 33 |
print(f"Error installing espeak-ng: {e}")
|
| 34 |
|
|
|
|
| 84 |
os.makedirs("wav", exist_ok=True)
|
| 85 |
os.makedirs("ckpts/Vevo", exist_ok=True)
|
| 86 |
|
| 87 |
+
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 88 |
+
|
| 89 |
+
# تابع ذخیره سازی امن (جایگزین torchaudio.save)
|
| 90 |
+
def my_save_audio(waveform, output_path, sample_rate=24000):
|
| 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 |
+
sf.write(output_path, waveform, sample_rate)
|
| 98 |
+
except Exception as e:
|
| 99 |
+
print(f"Save error: {e}")
|
| 100 |
+
raise e
|
| 101 |
|
| 102 |
def setup_configs():
|
| 103 |
if downloaded_resources["configs"]: return
|
|
|
|
| 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 |
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 |
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 |
|
| 214 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 215 |
|
| 216 |
+
# استفاده از soundfile برای ذخیره (چون torchaudio در نسخه جدید ارور میدهد)
|
| 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 ({content_tensor.shape[1]/24000:.2f}s)...")
|
| 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,
|
|
|
|
| 232 |
print("Warning: NaN fixed")
|
| 233 |
gen_audio = torch.nan_to_num(gen_audio, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 234 |
|
| 235 |
+
my_save_audio(gen_audio, output_path=output_path)
|
| 236 |
return output_path
|
| 237 |
|
| 238 |
finally:
|
|
|
|
| 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 |
|
| 243 |
+
with gr.Blocks(title="Vevo-Timbre (Secure)") as demo:
|
| 244 |
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 245 |
|
| 246 |
with gr.Row():
|