Update app.py
Browse files
app.py
CHANGED
|
@@ -174,7 +174,24 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 174 |
raise ValueError("Please upload audio files")
|
| 175 |
|
| 176 |
try:
|
| 177 |
-
# --- آماده سازی
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
if isinstance(reference_wav, tuple):
|
| 179 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 180 |
else:
|
|
@@ -188,43 +205,27 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 188 |
ref_tensor = torchaudio.functional.resample(ref_tensor, ref_sr, 24000)
|
| 189 |
ref_sr = 24000
|
| 190 |
|
| 191 |
-
#
|
| 192 |
-
|
| 193 |
-
ref_tensor = ref_tensor /
|
| 194 |
|
| 195 |
-
# برش رفرنس به 20 ثانیه
|
| 196 |
if ref_tensor.shape[1] > 24000 * 20:
|
| 197 |
ref_tensor = ref_tensor[:, :24000 * 20]
|
| 198 |
|
| 199 |
save_audio_pcm16(ref_tensor, temp_reference_path, ref_sr)
|
| 200 |
-
|
| 201 |
-
# --- آماده سازی Content ---
|
| 202 |
-
if isinstance(content_wav, tuple):
|
| 203 |
-
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 204 |
-
else:
|
| 205 |
-
content_sr, content_data = content_wav
|
| 206 |
-
|
| 207 |
-
if len(content_data.shape) > 1 and content_data.shape[1] > 1:
|
| 208 |
-
content_data = np.mean(content_data, axis=1)
|
| 209 |
|
| 210 |
-
|
| 211 |
-
if content_sr != 24000:
|
| 212 |
-
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 213 |
-
content_sr = 24000
|
| 214 |
-
|
| 215 |
-
# نرمال سازی هوشمند: صدای ورودی را همسطح صدای رفرنس میکنیم
|
| 216 |
-
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 217 |
-
|
| 218 |
-
# --- منطق Chunking ---
|
| 219 |
pipeline = get_pipeline()
|
| 220 |
|
| 221 |
SR = 24000
|
| 222 |
-
CHUNK_LEN = 10 * SR
|
| 223 |
-
|
|
|
|
|
|
|
| 224 |
INPUT_SIZE = CHUNK_LEN + OVERLAP
|
| 225 |
|
| 226 |
total_samples = content_tensor.shape[1]
|
| 227 |
-
print(f"[{session_id}] High Quality
|
| 228 |
|
| 229 |
final_parts = []
|
| 230 |
overlap_buffer = None
|
|
@@ -239,7 +240,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 239 |
gen = pipeline.inference_fm(
|
| 240 |
src_wav_path=temp_content_path,
|
| 241 |
timbre_ref_wav_path=temp_reference_path,
|
| 242 |
-
flow_matching_steps=64,
|
| 243 |
)
|
| 244 |
|
| 245 |
if torch.isnan(gen).any(): gen = torch.nan_to_num(gen, nan=0.0)
|
|
@@ -257,6 +258,7 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 257 |
head_to_mix = gen[:mix_len]
|
| 258 |
body_rest = gen[mix_len:]
|
| 259 |
|
|
|
|
| 260 |
alpha = np.linspace(0, 1, mix_len)
|
| 261 |
blended_segment = (overlap_buffer * (1 - alpha)) + (head_to_mix * alpha)
|
| 262 |
|
|
@@ -302,22 +304,15 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 302 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 303 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 304 |
|
| 305 |
-
with gr.Blocks(title="Vevo-Timbre (
|
| 306 |
-
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion
|
| 307 |
-
gr.Markdown(""
|
| 308 |
-
**ویژگیها:**
|
| 309 |
-
- **Steps 64:** کیفیت و دقت بافت صدا دو برابر شده است.
|
| 310 |
-
- **Auto-Leveling:** سطح صدای شما با مدل تنظیم میشود.
|
| 311 |
-
- **Seamless Stitching:** بدون پرش و بدون اضافه شدن زمان.
|
| 312 |
-
|
| 313 |
-
**نکته مهم:** برای بهترین نتیجه، سعی کنید **لحن، سرعت و احساس** صدای خودتان را شبیه فایل هدف کنید. مدل فقط جنس صدا را تغییر میدهد، نه بازیگری شما را!
|
| 314 |
-
""")
|
| 315 |
|
| 316 |
with gr.Row():
|
| 317 |
with gr.Column():
|
| 318 |
-
timbre_content = gr.Audio(label="Source Audio
|
| 319 |
-
timbre_reference = gr.Audio(label="Target Timbre
|
| 320 |
-
timbre_button = gr.Button("Generate
|
| 321 |
with gr.Column():
|
| 322 |
timbre_output = gr.Audio(label="Result")
|
| 323 |
|
|
|
|
| 174 |
raise ValueError("Please upload audio files")
|
| 175 |
|
| 176 |
try:
|
| 177 |
+
# --- آماده سازی Content ---
|
| 178 |
+
if isinstance(content_wav, tuple):
|
| 179 |
+
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
| 180 |
+
else:
|
| 181 |
+
content_sr, content_data = content_wav
|
| 182 |
+
|
| 183 |
+
if len(content_data.shape) > 1 and content_data.shape[1] > 1:
|
| 184 |
+
content_data = np.mean(content_data, axis=1)
|
| 185 |
+
|
| 186 |
+
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
| 187 |
+
if content_sr != 24000:
|
| 188 |
+
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, 24000)
|
| 189 |
+
content_sr = 24000
|
| 190 |
+
|
| 191 |
+
# نرمال سازی
|
| 192 |
+
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 193 |
+
|
| 194 |
+
# --- آماده سازی Reference ---
|
| 195 |
if isinstance(reference_wav, tuple):
|
| 196 |
ref_sr, ref_data = reference_wav if isinstance(reference_wav[0], int) else (reference_wav[1], reference_wav[0])
|
| 197 |
else:
|
|
|
|
| 205 |
ref_tensor = torchaudio.functional.resample(ref_tensor, ref_sr, 24000)
|
| 206 |
ref_sr = 24000
|
| 207 |
|
| 208 |
+
# تنظیم لول رفرنس
|
| 209 |
+
ref_max = torch.max(torch.abs(ref_tensor)) + 1e-6
|
| 210 |
+
ref_tensor = ref_tensor / ref_max * 0.95
|
| 211 |
|
|
|
|
| 212 |
if ref_tensor.shape[1] > 24000 * 20:
|
| 213 |
ref_tensor = ref_tensor[:, :24000 * 20]
|
| 214 |
|
| 215 |
save_audio_pcm16(ref_tensor, temp_reference_path, ref_sr)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 216 |
|
| 217 |
+
# --- منطق Chunking (اصلاح شده: همپوشانی کوتاه) ---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 218 |
pipeline = get_pipeline()
|
| 219 |
|
| 220 |
SR = 24000
|
| 221 |
+
CHUNK_LEN = 10 * SR # 10 ثانیه اصلی
|
| 222 |
+
# تغییر مهم: کاهش همپوشانی به 0.1 ثانیه (100 میلی ثانیه)
|
| 223 |
+
# این باعث میشود اکو از بین برود ولی اتصال همچنان نرم باشد
|
| 224 |
+
OVERLAP = int(0.1 * SR)
|
| 225 |
INPUT_SIZE = CHUNK_LEN + OVERLAP
|
| 226 |
|
| 227 |
total_samples = content_tensor.shape[1]
|
| 228 |
+
print(f"[{session_id}] Processing (High Quality 64 Steps)...")
|
| 229 |
|
| 230 |
final_parts = []
|
| 231 |
overlap_buffer = None
|
|
|
|
| 240 |
gen = pipeline.inference_fm(
|
| 241 |
src_wav_path=temp_content_path,
|
| 242 |
timbre_ref_wav_path=temp_reference_path,
|
| 243 |
+
flow_matching_steps=64, # کیفیت بالا
|
| 244 |
)
|
| 245 |
|
| 246 |
if torch.isnan(gen).any(): gen = torch.nan_to_num(gen, nan=0.0)
|
|
|
|
| 258 |
head_to_mix = gen[:mix_len]
|
| 259 |
body_rest = gen[mix_len:]
|
| 260 |
|
| 261 |
+
# میکس سریع (Fast Cross-Fade)
|
| 262 |
alpha = np.linspace(0, 1, mix_len)
|
| 263 |
blended_segment = (overlap_buffer * (1 - alpha)) + (head_to_mix * alpha)
|
| 264 |
|
|
|
|
| 304 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 305 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 306 |
|
| 307 |
+
with gr.Blocks(title="Vevo-Timbre (No Echo)") as demo:
|
| 308 |
+
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 309 |
+
gr.Markdown("نسخه اصلاح شده: حذف اکو در نقاط اتصال + کیفیت بالای ۶۴ مرحلهای.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 310 |
|
| 311 |
with gr.Row():
|
| 312 |
with gr.Column():
|
| 313 |
+
timbre_content = gr.Audio(label="Source Audio", type="numpy")
|
| 314 |
+
timbre_reference = gr.Audio(label="Target Timbre", type="numpy")
|
| 315 |
+
timbre_button = gr.Button("Generate", variant="primary")
|
| 316 |
with gr.Column():
|
| 317 |
timbre_output = gr.Audio(label="Result")
|
| 318 |
|