Update app.py
Browse files
app.py
CHANGED
|
@@ -11,9 +11,9 @@ from huggingface_hub import snapshot_download, hf_hub_download
|
|
| 11 |
import subprocess
|
| 12 |
import re
|
| 13 |
import spaces
|
| 14 |
-
import
|
| 15 |
|
| 16 |
-
# فقط منابع
|
| 17 |
downloaded_resources = {
|
| 18 |
"configs": False,
|
| 19 |
"tokenizer_vq8192": False,
|
|
@@ -22,7 +22,6 @@ downloaded_resources = {
|
|
| 22 |
}
|
| 23 |
|
| 24 |
def install_espeak():
|
| 25 |
-
"""Detect and install espeak-ng dependency"""
|
| 26 |
try:
|
| 27 |
result = subprocess.run(["which", "espeak-ng"], capture_output=True, text=True)
|
| 28 |
if result.returncode != 0:
|
|
@@ -30,7 +29,7 @@ def install_espeak():
|
|
| 30 |
subprocess.run(["apt-get", "update"], check=True)
|
| 31 |
subprocess.run(["apt-get", "install", "-y", "espeak-ng", "espeak-ng-data"], check=True)
|
| 32 |
else:
|
| 33 |
-
print("espeak-ng is
|
| 34 |
except Exception as e:
|
| 35 |
print(f"Error installing espeak-ng: {e}")
|
| 36 |
|
|
@@ -69,9 +68,7 @@ def patch_langsegment_init():
|
|
| 69 |
import LangSegment
|
| 70 |
importlib.reload(LangSegment)
|
| 71 |
except: pass
|
| 72 |
-
|
| 73 |
-
except Exception as e:
|
| 74 |
-
print(f"Error patching LangSegment: {e}")
|
| 75 |
|
| 76 |
patch_langsegment_init()
|
| 77 |
|
|
@@ -88,22 +85,8 @@ if os.path.dirname(os.path.abspath("Amphion")) not in sys.path:
|
|
| 88 |
os.makedirs("wav", exist_ok=True)
|
| 89 |
os.makedirs("ckpts/Vevo", exist_ok=True)
|
| 90 |
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
# تابع ذخیره سازی اختصاصی
|
| 94 |
-
def my_save_audio(waveform, output_path, sample_rate=24000):
|
| 95 |
-
try:
|
| 96 |
-
if isinstance(waveform, torch.Tensor):
|
| 97 |
-
waveform = waveform.detach().cpu()
|
| 98 |
-
if waveform.dim() == 2 and waveform.shape[0] == 1:
|
| 99 |
-
waveform = waveform.squeeze(0)
|
| 100 |
-
waveform = waveform.numpy()
|
| 101 |
-
|
| 102 |
-
sf.write(output_path, waveform, sample_rate)
|
| 103 |
-
print(f"Audio saved successfully to {output_path}")
|
| 104 |
-
except Exception as e:
|
| 105 |
-
print(f"Failed to save audio with soundfile: {e}")
|
| 106 |
-
raise e
|
| 107 |
|
| 108 |
def setup_configs():
|
| 109 |
if downloaded_resources["configs"]: return
|
|
@@ -128,7 +111,7 @@ print(f"Using device: {device}")
|
|
| 128 |
inference_pipelines = {}
|
| 129 |
|
| 130 |
def preload_all_resources():
|
| 131 |
-
print("Preloading
|
| 132 |
setup_configs()
|
| 133 |
|
| 134 |
global downloaded_content_style_tokenizer_path
|
|
@@ -149,8 +132,7 @@ def preload_all_resources():
|
|
| 149 |
local_dir = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"])
|
| 150 |
downloaded_vocoder_path = local_dir
|
| 151 |
downloaded_resources["vocoder"] = True
|
| 152 |
-
|
| 153 |
-
print("Timbre resources ready!")
|
| 154 |
|
| 155 |
downloaded_content_style_tokenizer_path = None
|
| 156 |
downloaded_fmt_path = None
|
|
@@ -162,18 +144,12 @@ def get_pipeline():
|
|
| 162 |
if "timbre" in inference_pipelines:
|
| 163 |
return inference_pipelines["timbre"]
|
| 164 |
|
| 165 |
-
content_style_tokenizer_ckpt_path = os.path.join(downloaded_content_style_tokenizer_path, "tokenizer/vq8192")
|
| 166 |
-
fmt_cfg_path = "./models/vc/vevo/config/Vq8192ToMels.json"
|
| 167 |
-
fmt_ckpt_path = os.path.join(downloaded_fmt_path, "acoustic_modeling/Vq8192ToMels")
|
| 168 |
-
vocoder_cfg_path = "./models/vc/vevo/config/Vocoder.json"
|
| 169 |
-
vocoder_ckpt_path = os.path.join(downloaded_vocoder_path, "acoustic_modeling/Vocoder")
|
| 170 |
-
|
| 171 |
pipeline = VevoInferencePipeline(
|
| 172 |
-
content_style_tokenizer_ckpt_path=
|
| 173 |
-
fmt_cfg_path=
|
| 174 |
-
fmt_ckpt_path=
|
| 175 |
-
vocoder_cfg_path=
|
| 176 |
-
vocoder_ckpt_path=
|
| 177 |
device=device,
|
| 178 |
)
|
| 179 |
|
|
@@ -182,122 +158,85 @@ def get_pipeline():
|
|
| 182 |
|
| 183 |
@spaces.GPU()
|
| 184 |
def vevo_timbre(content_wav, reference_wav):
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
|
|
|
|
|
|
| 188 |
|
| 189 |
if content_wav is None or reference_wav is None:
|
| 190 |
raise ValueError("Please upload audio files")
|
| 191 |
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
|
|
|
| 197 |
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
|
| 214 |
-
|
| 215 |
-
|
| 216 |
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 223 |
-
|
| 224 |
-
# برش زدن صدای رفرنس به 20 ثانیه اول (برای جلوگیری از گیج شدن مدل)
|
| 225 |
-
# صدای رفرنس فقط برای برداشتن "رنگ صدا" استفاده میشه و 20 ثانیه کافیه
|
| 226 |
-
if ref_tensor.shape[1] > 24000 * 20:
|
| 227 |
-
ref_tensor = ref_tensor[:, :24000 * 20]
|
| 228 |
-
|
| 229 |
-
# ذخیره موقت صدای رفرنس
|
| 230 |
-
sf.write(temp_reference_path, ref_tensor.squeeze().cpu().numpy(), ref_sr)
|
| 231 |
-
|
| 232 |
-
print(f"Total Duration: {content_tensor.shape[1]/24000:.2f}s")
|
| 233 |
-
|
| 234 |
-
# --- تکه تکه کردن صدای اصلی (Chunking Logic) ---
|
| 235 |
-
pipeline = get_pipeline()
|
| 236 |
-
|
| 237 |
-
CHUNK_DURATION = 15 # ثانیه (اندازه هر تکه)
|
| 238 |
-
CHUNK_SAMPLES = CHUNK_DURATION * 24000
|
| 239 |
-
total_samples = content_tensor.shape[1]
|
| 240 |
-
|
| 241 |
-
generated_chunks = []
|
| 242 |
-
|
| 243 |
-
# حلقه برای پردازش تکه تکه
|
| 244 |
-
for i in range(0, total_samples, CHUNK_SAMPLES):
|
| 245 |
-
end = min(i + CHUNK_SAMPLES, total_samples)
|
| 246 |
-
chunk = content_tensor[:, i:end]
|
| 247 |
|
| 248 |
-
|
| 249 |
|
| 250 |
-
# ذخیره
|
| 251 |
-
|
|
|
|
| 252 |
|
| 253 |
-
|
| 254 |
-
# پردازش تکه
|
| 255 |
-
gen_chunk = pipeline.inference_fm(
|
| 256 |
-
src_wav_path=temp_content_path,
|
| 257 |
-
timbre_ref_wav_path=temp_reference_path,
|
| 258 |
-
flow_matching_steps=32,
|
| 259 |
-
)
|
| 260 |
-
|
| 261 |
-
# بررسی خرابی احتمالی
|
| 262 |
-
if torch.isnan(gen_chunk).any() or torch.isinf(gen_chunk).any():
|
| 263 |
-
print("Warning: NaN in chunk, fixing...")
|
| 264 |
-
gen_chunk = torch.nan_to_num(gen_chunk, nan=0.0, posinf=0.95, neginf=-0.95)
|
| 265 |
-
|
| 266 |
-
# اضافه کردن به لیست خروجیها (��طمئن میشیم دوبعدی باشه [1, T])
|
| 267 |
-
if gen_chunk.dim() == 1:
|
| 268 |
-
gen_chunk = gen_chunk.unsqueeze(0)
|
| 269 |
-
generated_chunks.append(gen_chunk.cpu())
|
| 270 |
-
|
| 271 |
-
except Exception as e:
|
| 272 |
-
print(f"Error processing chunk starting at {i}: {e}")
|
| 273 |
-
# در صورت خطا در یک تکه، سکوت جایگزین میکنیم تا کل فایل خراب نشه
|
| 274 |
-
silence = torch.zeros_like(chunk)
|
| 275 |
-
generated_chunks.append(silence)
|
| 276 |
-
|
| 277 |
-
# --- چسباندن تکهها به هم ---
|
| 278 |
-
if not generated_chunks:
|
| 279 |
-
raise ValueError("No audio generated")
|
| 280 |
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 288 |
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion (Unlimited Length)")
|
| 292 |
-
gr.Markdown("این نسخه فایلهای طولانی را به صورت اتوماتیک به تکههای ۱۵ ثانیهای تقسیم کرده و پردازش میکند تا صدا خراب نشود.")
|
| 293 |
|
| 294 |
with gr.Row():
|
| 295 |
with gr.Column():
|
| 296 |
-
timbre_content = gr.Audio(label="Source Audio
|
| 297 |
-
timbre_reference = gr.Audio(label="Target Timbre
|
| 298 |
-
timbre_button = gr.Button("Generate
|
| 299 |
with gr.Column():
|
| 300 |
-
timbre_output = gr.Audio(label="Result
|
| 301 |
|
| 302 |
timbre_button.click(vevo_timbre, inputs=[timbre_content, timbre_reference], outputs=timbre_output)
|
| 303 |
|
|
|
|
| 11 |
import subprocess
|
| 12 |
import re
|
| 13 |
import spaces
|
| 14 |
+
import uuid
|
| 15 |
|
| 16 |
+
# دانلود فقط منابع ضروری
|
| 17 |
downloaded_resources = {
|
| 18 |
"configs": False,
|
| 19 |
"tokenizer_vq8192": False,
|
|
|
|
| 22 |
}
|
| 23 |
|
| 24 |
def install_espeak():
|
|
|
|
| 25 |
try:
|
| 26 |
result = subprocess.run(["which", "espeak-ng"], capture_output=True, text=True)
|
| 27 |
if result.returncode != 0:
|
|
|
|
| 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 |
|
|
|
|
| 68 |
import LangSegment
|
| 69 |
importlib.reload(LangSegment)
|
| 70 |
except: pass
|
| 71 |
+
except: pass
|
|
|
|
|
|
|
| 72 |
|
| 73 |
patch_langsegment_init()
|
| 74 |
|
|
|
|
| 85 |
os.makedirs("wav", exist_ok=True)
|
| 86 |
os.makedirs("ckpts/Vevo", exist_ok=True)
|
| 87 |
|
| 88 |
+
# اینجا دیگر مشکلی ندارد چون نسخه torchaudio را درست کردیم
|
| 89 |
+
from models.vc.vevo.vevo_utils import VevoInferencePipeline, save_audio
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
|
| 91 |
def setup_configs():
|
| 92 |
if downloaded_resources["configs"]: return
|
|
|
|
| 111 |
inference_pipelines = {}
|
| 112 |
|
| 113 |
def preload_all_resources():
|
| 114 |
+
print("Preloading resources...")
|
| 115 |
setup_configs()
|
| 116 |
|
| 117 |
global downloaded_content_style_tokenizer_path
|
|
|
|
| 132 |
local_dir = snapshot_download(repo_id="amphion/Vevo", repo_type="model", cache_dir="./ckpts/Vevo", allow_patterns=["acoustic_modeling/Vocoder/*"])
|
| 133 |
downloaded_vocoder_path = local_dir
|
| 134 |
downloaded_resources["vocoder"] = True
|
| 135 |
+
print("Resources ready.")
|
|
|
|
| 136 |
|
| 137 |
downloaded_content_style_tokenizer_path = None
|
| 138 |
downloaded_fmt_path = None
|
|
|
|
| 144 |
if "timbre" in inference_pipelines:
|
| 145 |
return inference_pipelines["timbre"]
|
| 146 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
pipeline = VevoInferencePipeline(
|
| 148 |
+
content_style_tokenizer_ckpt_path=os.path.join(downloaded_content_style_tokenizer_path, "tokenizer/vq8192"),
|
| 149 |
+
fmt_cfg_path="./models/vc/vevo/config/Vq8192ToMels.json",
|
| 150 |
+
fmt_ckpt_path=os.path.join(downloaded_fmt_path, "acoustic_modeling/Vq8192ToMels"),
|
| 151 |
+
vocoder_cfg_path="./models/vc/vevo/config/Vocoder.json",
|
| 152 |
+
vocoder_ckpt_path=os.path.join(downloaded_vocoder_path, "acoustic_modeling/Vocoder"),
|
| 153 |
device=device,
|
| 154 |
)
|
| 155 |
|
|
|
|
| 158 |
|
| 159 |
@spaces.GPU()
|
| 160 |
def vevo_timbre(content_wav, reference_wav):
|
| 161 |
+
# ایجاد نام فایل منحصر به فرد برای جلوگیری از تداخل کاربران
|
| 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"
|
| 165 |
+
output_path = f"wav/out_{session_id}.wav"
|
| 166 |
|
| 167 |
if content_wav is None or reference_wav is None:
|
| 168 |
raise ValueError("Please upload audio files")
|
| 169 |
|
| 170 |
+
try:
|
| 171 |
+
# --- پردازش صدای اصلی ---
|
| 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:
|
| 175 |
+
content_sr, content_data = content_wav
|
| 176 |
|
| 177 |
+
if len(content_data.shape) > 1 and content_data.shape[1] > 1:
|
| 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 صدای رفرنس ---
|
| 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:
|
| 191 |
+
ref_sr, ref_data = reference_wav
|
| 192 |
|
| 193 |
+
if len(ref_data.shape) > 1 and ref_data.shape[1] > 1:
|
| 194 |
+
ref_data = np.mean(ref_data, axis=1)
|
| 195 |
|
| 196 |
+
ref_tensor = torch.FloatTensor(ref_data).unsqueeze(0)
|
| 197 |
+
if ref_sr != 24000:
|
| 198 |
+
ref_tensor = torchaudio.functional.resample(ref_tensor, ref_sr, 24000)
|
| 199 |
+
ref_sr = 24000
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 200 |
|
| 201 |
+
ref_tensor = ref_tensor / (torch.max(torch.abs(ref_tensor)) + 1e-6) * 0.95
|
| 202 |
|
| 203 |
+
# ذخیره فایلها با torchaudio (چون نسخه قدیمی است، بدون ارور کار میکند و فرمت دقیق را حفظ میکند)
|
| 204 |
+
torchaudio.save(temp_content_path, content_tensor, content_sr)
|
| 205 |
+
torchaudio.save(temp_reference_path, ref_tensor, ref_sr)
|
| 206 |
|
| 207 |
+
print(f"[{session_id}] Processing Audio...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 208 |
|
| 209 |
+
pipeline = get_pipeline()
|
| 210 |
+
|
| 211 |
+
# اجرای مدل روی کل فایل (بدون تکه تکه کردن - چون قبلاً اینطوری کار میکرد)
|
| 212 |
+
gen_audio = pipeline.inference_fm(
|
| 213 |
+
src_wav_path=temp_content_path,
|
| 214 |
+
timbre_ref_wav_path=temp_reference_path,
|
| 215 |
+
flow_matching_steps=32,
|
| 216 |
+
)
|
| 217 |
+
|
| 218 |
+
if torch.isnan(gen_audio).any() or torch.isinf(gen_audio).any():
|
| 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 |
+
save_audio(gen_audio, output_path=output_path)
|
| 223 |
+
return output_path
|
| 224 |
+
|
| 225 |
+
finally:
|
| 226 |
+
# پاکسازی فایلهای موقت
|
| 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 (Stable)") as demo:
|
| 231 |
+
gr.Markdown("## Vevo-Timbre: Zero-Shot Voice Conversion")
|
|
|
|
|
|
|
| 232 |
|
| 233 |
with gr.Row():
|
| 234 |
with gr.Column():
|
| 235 |
+
timbre_content = gr.Audio(label="Source Audio", type="numpy")
|
| 236 |
+
timbre_reference = gr.Audio(label="Target Timbre", type="numpy")
|
| 237 |
+
timbre_button = gr.Button("Generate", variant="primary")
|
| 238 |
with gr.Column():
|
| 239 |
+
timbre_output = gr.Audio(label="Result")
|
| 240 |
|
| 241 |
timbre_button.click(vevo_timbre, inputs=[timbre_content, timbre_reference], outputs=timbre_output)
|
| 242 |
|