Update app.py
Browse files
app.py
CHANGED
|
@@ -13,11 +13,8 @@ import re
|
|
| 13 |
import spaces
|
| 14 |
import uuid
|
| 15 |
import soundfile as sf
|
| 16 |
-
# اضافه شدن کتابخانه PyDub
|
| 17 |
-
from pydub import AudioSegment
|
| 18 |
-
import io
|
| 19 |
|
| 20 |
-
# ---
|
| 21 |
downloaded_resources = {
|
| 22 |
"configs": False,
|
| 23 |
"tokenizer_vq8192": False,
|
|
@@ -75,22 +72,13 @@ os.makedirs("ckpts/Vevo", exist_ok=True)
|
|
| 75 |
|
| 76 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 77 |
|
| 78 |
-
# --- توابع کمکی جدید برای PyDub ---
|
| 79 |
-
def numpy_to_audiosegment(audio_arr, sample_rate=24000):
|
| 80 |
-
"""تبدیل آرایه نامپای (Float32) به آبجکت AudioSegment"""
|
| 81 |
-
# تبدیل به PCM 16-bit
|
| 82 |
-
audio_int16 = (audio_arr * 32767).astype(np.int16)
|
| 83 |
-
# ایجاد فایل در حافظه
|
| 84 |
-
byte_io = io.BytesIO()
|
| 85 |
-
sf.write(byte_io, audio_int16, sample_rate, format='WAV', subtype='PCM_16')
|
| 86 |
-
byte_io.seek(0)
|
| 87 |
-
return AudioSegment.from_wav(byte_io)
|
| 88 |
-
|
| 89 |
def save_audio_pcm16(waveform, output_path, sample_rate=24000):
|
| 90 |
-
# این تابع فقط برای ذخیره فایلهای موقت ورودی مدل است
|
| 91 |
try:
|
| 92 |
if isinstance(waveform, torch.Tensor):
|
| 93 |
-
waveform = waveform.detach().cpu()
|
|
|
|
|
|
|
|
|
|
| 94 |
sf.write(output_path, waveform, sample_rate, subtype='PCM_16')
|
| 95 |
except Exception as e:
|
| 96 |
print(f"Save error: {e}")
|
|
@@ -115,10 +103,8 @@ device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cp
|
|
| 115 |
inference_pipelines = {}
|
| 116 |
|
| 117 |
def preload_all_resources():
|
| 118 |
-
print("Preloading resources...")
|
| 119 |
setup_configs()
|
| 120 |
global downloaded_content_style_tokenizer_path, downloaded_fmt_path, downloaded_vocoder_path
|
| 121 |
-
|
| 122 |
if not downloaded_resources["tokenizer_vq8192"]:
|
| 123 |
downloaded_content_style_tokenizer_path = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"])
|
| 124 |
downloaded_resources["tokenizer_vq8192"] = True
|
|
@@ -128,7 +114,6 @@ def preload_all_resources():
|
|
| 128 |
if not downloaded_resources["vocoder"]:
|
| 129 |
downloaded_vocoder_path = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"])
|
| 130 |
downloaded_resources["vocoder"] = True
|
| 131 |
-
print("Resources ready.")
|
| 132 |
|
| 133 |
downloaded_content_style_tokenizer_path = None
|
| 134 |
downloaded_fmt_path = None
|
|
@@ -159,6 +144,8 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 159 |
raise ValueError("Please upload audio files")
|
| 160 |
|
| 161 |
try:
|
|
|
|
|
|
|
| 162 |
# --- 1. پردازش ورودی ---
|
| 163 |
if isinstance(content_wav, tuple):
|
| 164 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
|
@@ -167,8 +154,8 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 167 |
if len(content_data.shape) > 1: content_data = np.mean(content_data, axis=1)
|
| 168 |
|
| 169 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
| 170 |
-
if content_sr !=
|
| 171 |
-
content_tensor = torchaudio.functional.resample(content_tensor, content_sr,
|
| 172 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 173 |
content_full_np = content_tensor.squeeze().numpy()
|
| 174 |
|
|
@@ -180,37 +167,56 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 180 |
if len(ref_data.shape) > 1: ref_data = np.mean(ref_data, axis=1)
|
| 181 |
|
| 182 |
ref_tensor = torch.FloatTensor(ref_data).unsqueeze(0)
|
| 183 |
-
if ref_sr !=
|
| 184 |
-
ref_tensor = torchaudio.functional.resample(ref_tensor, ref_sr,
|
| 185 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 186 |
-
if ref_tensor.shape[1] >
|
| 187 |
-
save_audio_pcm16(ref_tensor, temp_reference_path,
|
| 188 |
|
| 189 |
-
# --- 3.
|
| 190 |
pipeline = get_pipeline()
|
| 191 |
-
SR = 24000
|
| 192 |
|
| 193 |
-
|
| 194 |
-
|
|
|
|
| 195 |
|
| 196 |
-
|
| 197 |
-
|
| 198 |
total_samples = len(content_full_np)
|
| 199 |
|
| 200 |
-
|
| 201 |
-
final_audio_segment = AudioSegment.empty()
|
| 202 |
|
| 203 |
-
|
| 204 |
-
|
|
|
|
|
|
|
| 205 |
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 209 |
|
| 210 |
-
|
|
|
|
|
|
|
| 211 |
|
| 212 |
-
|
| 213 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 214 |
|
| 215 |
try:
|
| 216 |
gen = pipeline.inference_fm(
|
|
@@ -218,53 +224,63 @@ def vevo_timbre(content_wav, reference_wav):
|
|
| 218 |
timbre_ref_wav_path=temp_reference_path,
|
| 219 |
flow_matching_steps=64,
|
| 220 |
)
|
| 221 |
-
|
| 222 |
if torch.isnan(gen).any(): gen = torch.nan_to_num(gen, nan=0.0)
|
| 223 |
gen_np = gen.detach().cpu().squeeze().numpy()
|
| 224 |
|
| 225 |
-
#
|
| 226 |
-
trim_samples = current_cursor - start_slice
|
| 227 |
|
| 228 |
-
if
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
|
|
|
|
|
|
| 233 |
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
if len(final_audio_segment) > 0:
|
| 238 |
-
# تکنیک: یک فید بسیار ریز (Crossfade 5ms)
|
| 239 |
-
# نکته: PyDub برای کراسفید نیاز به همپوشانی دارد، اما چون ما کانتکست را دقیق بریدیم،
|
| 240 |
-
# اینجا از append ساده استفاده میکنیم و فقط لبهها را نرم میکنیم.
|
| 241 |
-
|
| 242 |
-
# نرم کردن ابتدای تکه جدید (Fade In 5ms)
|
| 243 |
-
new_segment = new_segment.fade_in(5)
|
| 244 |
-
# نرم کردن انتهای تکه قبلی (Fade Out 5ms) - (قبلاً انجام شده یا الان انجام میدیم)
|
| 245 |
-
# در اینجا فقط چسباندن (Append) با فید این کافیست.
|
| 246 |
-
|
| 247 |
-
final_audio_segment += new_segment
|
| 248 |
else:
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 253 |
except Exception as e:
|
| 254 |
-
print(f"Error: {e}")
|
| 255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 256 |
|
| 257 |
-
|
| 258 |
-
final_audio_segment.export(output_path, format="wav")
|
| 259 |
return output_path
|
| 260 |
|
| 261 |
finally:
|
| 262 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 263 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 264 |
|
| 265 |
-
with gr.Blocks(title="Vevo-Timbre (
|
| 266 |
-
gr.Markdown("## Vevo-Timbre: Voice Conversion")
|
| 267 |
-
gr.Markdown("
|
| 268 |
|
| 269 |
with gr.Row():
|
| 270 |
with gr.Column():
|
|
|
|
| 13 |
import spaces
|
| 14 |
import uuid
|
| 15 |
import soundfile as sf
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
+
# --- تنظیمات و نصب ---
|
| 18 |
downloaded_resources = {
|
| 19 |
"configs": False,
|
| 20 |
"tokenizer_vq8192": False,
|
|
|
|
| 72 |
|
| 73 |
from models.vc.vevo.vevo_utils import VevoInferencePipeline
|
| 74 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
def save_audio_pcm16(waveform, output_path, sample_rate=24000):
|
|
|
|
| 76 |
try:
|
| 77 |
if isinstance(waveform, torch.Tensor):
|
| 78 |
+
waveform = waveform.detach().cpu()
|
| 79 |
+
if waveform.dim() == 2 and waveform.shape[0] == 1:
|
| 80 |
+
waveform = waveform.squeeze(0)
|
| 81 |
+
waveform = waveform.numpy()
|
| 82 |
sf.write(output_path, waveform, sample_rate, subtype='PCM_16')
|
| 83 |
except Exception as e:
|
| 84 |
print(f"Save error: {e}")
|
|
|
|
| 103 |
inference_pipelines = {}
|
| 104 |
|
| 105 |
def preload_all_resources():
|
|
|
|
| 106 |
setup_configs()
|
| 107 |
global downloaded_content_style_tokenizer_path, downloaded_fmt_path, downloaded_vocoder_path
|
|
|
|
| 108 |
if not downloaded_resources["tokenizer_vq8192"]:
|
| 109 |
downloaded_content_style_tokenizer_path = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["tokenizer/vq8192/*"])
|
| 110 |
downloaded_resources["tokenizer_vq8192"] = True
|
|
|
|
| 114 |
if not downloaded_resources["vocoder"]:
|
| 115 |
downloaded_vocoder_path = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"])
|
| 116 |
downloaded_resources["vocoder"] = True
|
|
|
|
| 117 |
|
| 118 |
downloaded_content_style_tokenizer_path = None
|
| 119 |
downloaded_fmt_path = None
|
|
|
|
| 144 |
raise ValueError("Please upload audio files")
|
| 145 |
|
| 146 |
try:
|
| 147 |
+
SR = 24000
|
| 148 |
+
|
| 149 |
# --- 1. پردازش ورودی ---
|
| 150 |
if isinstance(content_wav, tuple):
|
| 151 |
content_sr, content_data = content_wav if isinstance(content_wav[0], int) else (content_wav[1], content_wav[0])
|
|
|
|
| 154 |
if len(content_data.shape) > 1: content_data = np.mean(content_data, axis=1)
|
| 155 |
|
| 156 |
content_tensor = torch.FloatTensor(content_data).unsqueeze(0)
|
| 157 |
+
if content_sr != SR:
|
| 158 |
+
content_tensor = torchaudio.functional.resample(content_tensor, content_sr, SR)
|
| 159 |
content_tensor = content_tensor / (torch.max(torch.abs(content_tensor)) + 1e-6) * 0.95
|
| 160 |
content_full_np = content_tensor.squeeze().numpy()
|
| 161 |
|
|
|
|
| 167 |
if len(ref_data.shape) > 1: ref_data = np.mean(ref_data, axis=1)
|
| 168 |
|
| 169 |
ref_tensor = torch.FloatTensor(ref_data).unsqueeze(0)
|
| 170 |
+
if ref_sr != SR:
|
| 171 |
+
ref_tensor = torchaudio.functional.resample(ref_tensor, ref_sr, SR)
|
| 172 |
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 173 |
+
if ref_tensor.shape[1] > SR * 20: ref_tensor = ref_tensor[:, :SR * 20]
|
| 174 |
+
save_audio_pcm16(ref_tensor, temp_reference_path, SR)
|
| 175 |
|
| 176 |
+
# --- 3. استراتژی جوش دادن Equal Power (500ms) ---
|
| 177 |
pipeline = get_pipeline()
|
|
|
|
| 178 |
|
| 179 |
+
# تنظیمات حیاتی
|
| 180 |
+
CHUNK_DURATION = 10.0 # طول خالص هر تکه
|
| 181 |
+
CROSSFADE_SEC = 0.5 # طول همپوشانی (نیم ثانیه برای حذف لرزش)
|
| 182 |
|
| 183 |
+
chunk_samples = int(CHUNK_DURATION * SR)
|
| 184 |
+
crossfade_samples = int(CROSSFADE_SEC * SR)
|
| 185 |
total_samples = len(content_full_np)
|
| 186 |
|
| 187 |
+
final_output = np.array([], dtype=np.float32)
|
|
|
|
| 188 |
|
| 189 |
+
# ایجاد منحنی فید Equal Power (سینوسی)
|
| 190 |
+
# این منحنی باعث میشود حجم صدا در محل اتصال ثابت بماند
|
| 191 |
+
fade_out_curve = np.cos(np.linspace(0, np.pi/2, crossfade_samples))
|
| 192 |
+
fade_in_curve = np.sin(np.linspace(0, np.pi/2, crossfade_samples))
|
| 193 |
|
| 194 |
+
# شروع حلقه پردازش
|
| 195 |
+
# ما در هر مرحله به اندازه chunk_samples جلو میرویم
|
| 196 |
+
# اما برای ورودی مدل، crossfade_samples را از قبل هم برمیداریم
|
| 197 |
+
|
| 198 |
+
cursor = 0
|
| 199 |
+
print(f"[{session_id}] Processing with 500ms Equal-Power Crossfade...")
|
| 200 |
+
|
| 201 |
+
while cursor < total_samples:
|
| 202 |
+
# تعیین بازه ورودی برای مدل
|
| 203 |
+
# اگر اولین تکه نیست، باید کمی از عقبتر شروع کنیم (برای همپوشانی)
|
| 204 |
+
is_first_chunk = (cursor == 0)
|
| 205 |
|
| 206 |
+
start_idx = cursor
|
| 207 |
+
if not is_first_chunk:
|
| 208 |
+
start_idx -= crossfade_samples # عقبگرد برای همپوشانی
|
| 209 |
|
| 210 |
+
end_idx = min(total_samples, cursor + chunk_samples)
|
| 211 |
+
|
| 212 |
+
# اگر به انتهای فایل رسیدیم و تکه خیلی کوچک است
|
| 213 |
+
if start_idx >= end_idx:
|
| 214 |
+
break
|
| 215 |
+
|
| 216 |
+
current_chunk_input = content_full_np[start_idx:end_idx]
|
| 217 |
+
|
| 218 |
+
# ذخیره و اجرا
|
| 219 |
+
save_audio_pcm16(torch.FloatTensor(current_chunk_input).unsqueeze(0), temp_content_path, SR)
|
| 220 |
|
| 221 |
try:
|
| 222 |
gen = pipeline.inference_fm(
|
|
|
|
| 224 |
timbre_ref_wav_path=temp_reference_path,
|
| 225 |
flow_matching_steps=64,
|
| 226 |
)
|
|
|
|
| 227 |
if torch.isnan(gen).any(): gen = torch.nan_to_num(gen, nan=0.0)
|
| 228 |
gen_np = gen.detach().cpu().squeeze().numpy()
|
| 229 |
|
| 230 |
+
# --- عملیات میکس هوشمند ---
|
|
|
|
| 231 |
|
| 232 |
+
if is_first_chunk:
|
| 233 |
+
# تکه اول: مستقیماً اضافه کن
|
| 234 |
+
final_output = np.concatenate([final_output, gen_np])
|
| 235 |
+
else:
|
| 236 |
+
# تکههای بعدی:
|
| 237 |
+
# 1. بخش همپوشانی (Crossfade Area)
|
| 238 |
+
# 2. بخش جدید (New Area)
|
| 239 |
|
| 240 |
+
if len(gen_np) < crossfade_samples:
|
| 241 |
+
# اگر خروجی خیلی کوتاه بود (نادر)، فقط بچسبان
|
| 242 |
+
final_output = np.concatenate([final_output, gen_np])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
else:
|
| 244 |
+
# جدا کردن بخش میکس و بخش جدید از خروجی فعلی
|
| 245 |
+
overlap_part_new = gen_np[:crossfade_samples]
|
| 246 |
+
rest_part_new = gen_np[crossfade_samples:]
|
| 247 |
+
|
| 248 |
+
# جدا کردن بخش میکس از انتهای خروجی قبلی
|
| 249 |
+
if len(final_output) >= crossfade_samples:
|
| 250 |
+
overlap_part_old = final_output[-crossfade_samples:]
|
| 251 |
+
|
| 252 |
+
# فرمول Equal Power Crossfade
|
| 253 |
+
# Old * Cos + New * Sin
|
| 254 |
+
blended = (overlap_part_old * fade_out_curve) + (overlap_part_new * fade_in_curve)
|
| 255 |
+
|
| 256 |
+
# جایگزینی انتهای آرایه اصلی با بخش میکس شده
|
| 257 |
+
final_output[-crossfade_samples:] = blended
|
| 258 |
+
|
| 259 |
+
# اضافه کردن باقیمانده
|
| 260 |
+
final_output = np.concatenate([final_output, rest_part_new])
|
| 261 |
+
else:
|
| 262 |
+
# اگر بافر قبلی خیلی کوتاه بود (نباید پیش بیاید)
|
| 263 |
+
final_output = np.concatenate([final_output, gen_np])
|
| 264 |
+
|
| 265 |
except Exception as e:
|
| 266 |
+
print(f"Error at {cursor}: {e}")
|
| 267 |
+
# در صورت خطا سکوت اضافه کن
|
| 268 |
+
missing = end_idx - start_idx
|
| 269 |
+
final_output = np.concatenate([final_output, np.zeros(missing)])
|
| 270 |
+
|
| 271 |
+
# حرکت به جلو
|
| 272 |
+
cursor += chunk_samples
|
| 273 |
|
| 274 |
+
save_audio_pcm16(final_output, output_path, SR)
|
|
|
|
| 275 |
return output_path
|
| 276 |
|
| 277 |
finally:
|
| 278 |
if os.path.exists(temp_content_path): os.remove(temp_content_path)
|
| 279 |
if os.path.exists(temp_reference_path): os.remove(temp_reference_path)
|
| 280 |
|
| 281 |
+
with gr.Blocks(title="Vevo-Timbre (Pro Stitch)") as demo:
|
| 282 |
+
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
| 283 |
+
gr.Markdown("Professional Stitching: 500ms Equal-Power Crossfade (No Jitter, No Ghosting).")
|
| 284 |
|
| 285 |
with gr.Row():
|
| 286 |
with gr.Column():
|