Spaces:
Runtime error
Runtime error
Upload 3 files
Browse files
app.py
CHANGED
|
@@ -3,16 +3,15 @@ import gradio as gr
|
|
| 3 |
import numpy as np
|
| 4 |
import soundfile as sf
|
| 5 |
import librosa
|
| 6 |
-
import torch
|
| 7 |
|
| 8 |
# 检查 GPU
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
-
|
| 13 |
-
SILERO_MODEL = None
|
| 14 |
-
SILERO_LOAD_ATTEMPTED = False
|
| 15 |
-
SILERO_LOAD_STATUS = "未尝试" # "未尝试", "加载中", "成功", "失败"
|
| 16 |
|
| 17 |
def extract_audio_from_video(video_path, output_path):
|
| 18 |
"""从视频中提取音频"""
|
|
@@ -98,153 +97,36 @@ def run_demucs_separation(audio_path, output_dir):
|
|
| 98 |
raise RuntimeError(f"Demucs 分离失败: {str(e)}")
|
| 99 |
|
| 100 |
|
| 101 |
-
def
|
| 102 |
"""
|
| 103 |
-
|
| 104 |
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
| 108 |
|
| 109 |
-
|
| 110 |
-
- 如果 Silero VAD 加载失败(超时或其他错误)
|
| 111 |
-
- 自动切换到传统多特征算法
|
| 112 |
-
- 准确率从 85-90% 降到 75-80%
|
| 113 |
"""
|
| 114 |
-
global SILERO_MODEL, SILERO_LOAD_ATTEMPTED, SILERO_LOAD_STATUS
|
| 115 |
-
|
| 116 |
-
# 如果已经尝试过加载,直接返回结果
|
| 117 |
-
if SILERO_LOAD_ATTEMPTED:
|
| 118 |
-
return SILERO_MODEL is not None
|
| 119 |
-
|
| 120 |
-
SILERO_LOAD_ATTEMPTED = True
|
| 121 |
-
SILERO_LOAD_STATUS = "加载中"
|
| 122 |
-
|
| 123 |
-
try:
|
| 124 |
-
print("📥 开始下载 Silero VAD 模型(3分钟超时)...")
|
| 125 |
-
|
| 126 |
-
# 使用 subprocess 控制超时
|
| 127 |
-
import signal
|
| 128 |
-
|
| 129 |
-
def timeout_handler(signum, frame):
|
| 130 |
-
raise TimeoutError("Silero 模型下载超时(3分钟限制)")
|
| 131 |
-
|
| 132 |
-
# 设置超时(只在 Linux/Mac 上有效)
|
| 133 |
-
if hasattr(signal, 'SIGALRM'):
|
| 134 |
-
signal.signal(signal.SIGALRM, timeout_handler)
|
| 135 |
-
signal.alarm(timeout)
|
| 136 |
-
|
| 137 |
-
try:
|
| 138 |
-
# 尝试从 torch.hub 加载
|
| 139 |
-
SILERO_MODEL, utils = torch.hub.load(
|
| 140 |
-
repo_or_dir='snakers4/silero-vad',
|
| 141 |
-
model='silero_vad',
|
| 142 |
-
force_reload=False,
|
| 143 |
-
onnx=False,
|
| 144 |
-
verbose=False
|
| 145 |
-
)
|
| 146 |
-
SILERO_MODEL = SILERO_MODEL.to(DEVICE)
|
| 147 |
-
SILERO_MODEL.eval()
|
| 148 |
-
|
| 149 |
-
SILERO_LOAD_STATUS = "成功"
|
| 150 |
-
print("✅ Silero VAD 模型加载成功")
|
| 151 |
-
return True
|
| 152 |
-
|
| 153 |
-
finally:
|
| 154 |
-
# 取消超时
|
| 155 |
-
if hasattr(signal, 'SIGALRM'):
|
| 156 |
-
signal.alarm(0)
|
| 157 |
-
|
| 158 |
-
except TimeoutError as e:
|
| 159 |
-
SILERO_LOAD_STATUS = "失败(超时)"
|
| 160 |
-
print(f"⚠️ {str(e)}")
|
| 161 |
-
print(" 【降级】自动切换到传统算法")
|
| 162 |
-
SILERO_MODEL = None
|
| 163 |
-
return False
|
| 164 |
-
|
| 165 |
-
except Exception as e:
|
| 166 |
-
SILERO_LOAD_STATUS = "失败(错误)"
|
| 167 |
-
print(f"⚠️ Silero VAD 加载失败: {str(e)}")
|
| 168 |
-
print(" 【降级】自动切换到传统算法")
|
| 169 |
-
SILERO_MODEL = None
|
| 170 |
-
return False
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
def detect_speech_with_silero(vocals_audio, sr):
|
| 174 |
-
"""使用 Silero VAD 深度学习模型检测说话"""
|
| 175 |
-
try:
|
| 176 |
-
global SILERO_MODEL
|
| 177 |
-
if SILERO_MODEL is None:
|
| 178 |
-
raise RuntimeError("Silero 模型未加载")
|
| 179 |
-
|
| 180 |
-
# 重采样到 16kHz
|
| 181 |
-
if sr != 16000:
|
| 182 |
-
vocals_16k = librosa.resample(vocals_audio, orig_sr=sr, target_sr=16000)
|
| 183 |
-
sr_work = 16000
|
| 184 |
-
else:
|
| 185 |
-
vocals_16k = vocals_audio
|
| 186 |
-
sr_work = 16000
|
| 187 |
-
|
| 188 |
-
# 转换为 torch tensor
|
| 189 |
-
audio_tensor = torch.from_numpy(vocals_16k).float().to(DEVICE)
|
| 190 |
-
|
| 191 |
-
# 使用 Silero VAD 检测
|
| 192 |
-
window_size_samples = 512
|
| 193 |
-
speech_probs = []
|
| 194 |
-
|
| 195 |
-
with torch.no_grad():
|
| 196 |
-
for i in range(0, len(audio_tensor), window_size_samples):
|
| 197 |
-
chunk = audio_tensor[i:i+window_size_samples]
|
| 198 |
-
if len(chunk) < window_size_samples:
|
| 199 |
-
chunk = torch.nn.functional.pad(chunk, (0, window_size_samples - len(chunk)))
|
| 200 |
-
|
| 201 |
-
speech_prob = SILERO_MODEL(chunk.unsqueeze(0), sr_work).item()
|
| 202 |
-
speech_probs.append(speech_prob)
|
| 203 |
-
|
| 204 |
-
# 创建掩码
|
| 205 |
-
speech_mask = np.repeat(speech_probs, window_size_samples)[:len(vocals_16k)]
|
| 206 |
-
speech_mask = (speech_mask > 0.5).astype(np.float32)
|
| 207 |
-
|
| 208 |
-
# 调整回原始采样率
|
| 209 |
-
if sr != sr_work:
|
| 210 |
-
from scipy.interpolate import interp1d
|
| 211 |
-
old_indices = np.linspace(0, 1, len(speech_mask))
|
| 212 |
-
new_indices = np.linspace(0, 1, len(vocals_audio))
|
| 213 |
-
interpolator = interp1d(old_indices, speech_mask, kind='linear', fill_value='extrapolate')
|
| 214 |
-
speech_mask = interpolator(new_indices)
|
| 215 |
-
|
| 216 |
-
# 确保长度匹配
|
| 217 |
-
if len(speech_mask) != len(vocals_audio):
|
| 218 |
-
if len(speech_mask) < len(vocals_audio):
|
| 219 |
-
speech_mask = np.pad(speech_mask, (0, len(vocals_audio) - len(speech_mask)))
|
| 220 |
-
else:
|
| 221 |
-
speech_mask = speech_mask[:len(vocals_audio)]
|
| 222 |
-
|
| 223 |
-
speech_mask = (speech_mask > 0.5).astype(np.float32)
|
| 224 |
-
|
| 225 |
-
return speech_mask
|
| 226 |
-
|
| 227 |
-
except Exception as e:
|
| 228 |
-
print(f"Silero VAD 检测失败: {str(e)}")
|
| 229 |
-
return None
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
def detect_speech_fallback(vocals_audio, sr):
|
| 233 |
-
"""传统算法备用方案(降级后使用)"""
|
| 234 |
try:
|
| 235 |
hop_length = 512
|
| 236 |
frame_length = 2048
|
| 237 |
|
| 238 |
-
# 能量
|
| 239 |
rms = librosa.feature.rms(y=vocals_audio, frame_length=frame_length, hop_length=hop_length)[0]
|
| 240 |
|
| 241 |
-
# 零交叉率
|
| 242 |
zcr = librosa.feature.zero_crossing_rate(vocals_audio, frame_length=frame_length, hop_length=hop_length)[0]
|
| 243 |
|
| 244 |
-
# 频谱质心
|
| 245 |
spectral_centroids = librosa.feature.spectral_centroid(y=vocals_audio, sr=sr, hop_length=hop_length)[0]
|
| 246 |
|
| 247 |
-
#
|
|
|
|
|
|
|
|
|
|
| 248 |
try:
|
| 249 |
f0, voiced_flag, voiced_probs = librosa.pyin(
|
| 250 |
vocals_audio,
|
|
@@ -255,26 +137,35 @@ def detect_speech_fallback(vocals_audio, sr):
|
|
| 255 |
hop_length=hop_length
|
| 256 |
)
|
| 257 |
f0 = np.nan_to_num(f0, nan=0.0)
|
|
|
|
| 258 |
except:
|
| 259 |
f0 = np.zeros(len(rms))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 260 |
|
| 261 |
-
# 归一化
|
| 262 |
-
min_len = min(len(rms), len(zcr), len(spectral_centroids), len(f0))
|
| 263 |
rms = rms[:min_len]
|
| 264 |
zcr = zcr[:min_len]
|
| 265 |
spectral_centroids = spectral_centroids[:min_len]
|
|
|
|
|
|
|
| 266 |
f0 = f0[:min_len]
|
| 267 |
|
| 268 |
# 说话特征得分
|
|
|
|
| 269 |
zcr_score = np.clip((zcr - 0.05) / 0.15, 0, 1)
|
| 270 |
|
|
|
|
| 271 |
rms_norm = rms / (np.max(rms) + 1e-8)
|
| 272 |
energy_variation = np.abs(np.gradient(rms_norm))
|
| 273 |
energy_score = np.clip(energy_variation * 10, 0, 1)
|
| 274 |
|
|
|
|
| 275 |
centroid_variation = np.abs(np.gradient(spectral_centroids))
|
| 276 |
centroid_score = np.clip(centroid_variation / (np.mean(centroid_variation) + 1e-8), 0, 1)
|
| 277 |
|
|
|
|
| 278 |
pitch_continuity = np.zeros_like(f0)
|
| 279 |
for i in range(1, len(f0)):
|
| 280 |
if f0[i] > 0 and f0[i-1] > 0:
|
|
@@ -290,9 +181,12 @@ def detect_speech_fallback(vocals_audio, sr):
|
|
| 290 |
0.20 * pitch_continuity
|
| 291 |
)
|
| 292 |
|
| 293 |
-
|
|
|
|
|
|
|
| 294 |
|
| 295 |
-
# 后处理
|
|
|
|
| 296 |
min_duration = int(0.2 * sr / hop_length)
|
| 297 |
i = 0
|
| 298 |
while i < len(speaking_mask):
|
|
@@ -306,15 +200,30 @@ def detect_speech_fallback(vocals_audio, sr):
|
|
| 306 |
else:
|
| 307 |
i += 1
|
| 308 |
|
| 309 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 310 |
speaking_mask_samples = np.repeat(speaking_mask, hop_length)
|
| 311 |
|
|
|
|
| 312 |
if len(speaking_mask_samples) < len(vocals_audio):
|
| 313 |
speaking_mask_samples = np.pad(speaking_mask_samples, (0, len(vocals_audio) - len(speaking_mask_samples)))
|
| 314 |
else:
|
| 315 |
speaking_mask_samples = speaking_mask_samples[:len(vocals_audio)]
|
| 316 |
|
| 317 |
-
# 平滑
|
| 318 |
smooth_window = int(0.03 * sr)
|
| 319 |
if smooth_window > 1:
|
| 320 |
speaking_mask_samples = np.convolve(
|
|
@@ -327,42 +236,14 @@ def detect_speech_fallback(vocals_audio, sr):
|
|
| 327 |
return speaking_mask_samples
|
| 328 |
|
| 329 |
except Exception as e:
|
| 330 |
-
print(f"
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
"""混合检测策略:优先使用 Silero VAD,失败则降级"""
|
| 336 |
-
try:
|
| 337 |
-
# 尝试加载 Silero 模型(懒加载,3分钟超时)
|
| 338 |
-
silero_available = load_silero_model_lazy(timeout=180)
|
| 339 |
-
|
| 340 |
-
if silero_available:
|
| 341 |
-
print("✅ 使用 Silero VAD 深度学习模型检测")
|
| 342 |
-
speech_mask = detect_speech_with_silero(vocals_audio, sr)
|
| 343 |
-
|
| 344 |
-
if speech_mask is not None:
|
| 345 |
-
if mode == 'strict':
|
| 346 |
-
from scipy.ndimage import binary_erosion
|
| 347 |
-
kernel_size = int(0.05 * sr)
|
| 348 |
-
if kernel_size > 1:
|
| 349 |
-
speech_mask = binary_erosion(speech_mask, structure=np.ones(kernel_size)).astype(np.float32)
|
| 350 |
-
|
| 351 |
-
singing_mask = 1 - speech_mask
|
| 352 |
-
return singing_mask, "Silero VAD"
|
| 353 |
-
|
| 354 |
-
# Silero 失败,降级到传统算法
|
| 355 |
-
print("⚠️ 【降级】使用传统多特征算法")
|
| 356 |
-
speech_mask = detect_speech_fallback(vocals_audio, sr)
|
| 357 |
-
singing_mask = 1 - speech_mask
|
| 358 |
-
return singing_mask, "传统算法"
|
| 359 |
-
|
| 360 |
-
except Exception as e:
|
| 361 |
-
print(f"检测失败: {str(e)}")
|
| 362 |
-
return np.ones(len(vocals_audio), dtype=np.float32), "传统算法"
|
| 363 |
|
| 364 |
|
| 365 |
-
def process_audio_full(audio_file,
|
| 366 |
"""完整的音频分离流程"""
|
| 367 |
if audio_file is None:
|
| 368 |
return None, None, None, "❌ 请先上传音频或视频文件"
|
|
@@ -388,6 +269,7 @@ def process_audio_full(audio_file, detection_mode, enable_detection):
|
|
| 388 |
save_audio(temp_wav, audio, sr)
|
| 389 |
|
| 390 |
# 2. Demucs 分离
|
|
|
|
| 391 |
status_messages.append("🎵 使用 Demucs AI 模型分离人声和伴奏...")
|
| 392 |
status_messages.append(" (首次运行会下载模型,约500MB)")
|
| 393 |
yield None, None, None, "\n".join(status_messages)
|
|
@@ -397,53 +279,33 @@ def process_audio_full(audio_file, detection_mode, enable_detection):
|
|
| 397 |
vocals, _ = librosa.load(vocals_path, sr=sr, mono=True)
|
| 398 |
instrumental, _ = librosa.load(instrumental_path, sr=sr, mono=True)
|
| 399 |
|
| 400 |
-
|
| 401 |
-
|
| 402 |
|
|
|
|
| 403 |
if enable_detection:
|
| 404 |
-
status_messages.append("
|
| 405 |
-
status_messages.append("
|
| 406 |
-
status_messages.append("
|
| 407 |
-
status_messages.append("
|
| 408 |
-
status_messages.append(" 如超时将自动【降级】到传统算法")
|
| 409 |
yield None, None, None, "\n".join(status_messages)
|
| 410 |
|
| 411 |
-
#
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
| 415 |
|
| 416 |
-
|
| 417 |
-
global SILERO_LOAD_STATUS
|
| 418 |
-
if algorithm_used == "Silero VAD":
|
| 419 |
-
status_messages.append("✅✅✅ 检测器状态: Silero VAD 深度学习")
|
| 420 |
-
status_messages.append(" 📈 预期准确率: 85-90%")
|
| 421 |
-
status_messages.append(" 🎯 算法类型: 神经网络")
|
| 422 |
-
else:
|
| 423 |
-
status_messages.append("⚠️⚠️⚠️ 检测器状态: 传统多特征算法(已降级)")
|
| 424 |
-
status_messages.append(f" 🔴 降级原因: {SILERO_LOAD_STATUS}")
|
| 425 |
-
status_messages.append(" 📉 预期准确率: 75-80%")
|
| 426 |
-
status_messages.append(" 🎯 算法类型: 信号处理")
|
| 427 |
-
status_messages.append("")
|
| 428 |
-
status_messages.append(" 💡 提示: 如需高准确率,建议:")
|
| 429 |
-
status_messages.append(" 1. 刷新页面重试")
|
| 430 |
-
status_messages.append(" 2. 或使用稳定版(移除 Silero)")
|
| 431 |
-
|
| 432 |
-
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
| 433 |
-
status_messages.append("🎤 正在分析音频特征...")
|
| 434 |
-
yield None, None, None, "\n".join(status_messages)
|
| 435 |
else:
|
| 436 |
status_messages.append("⚠️ 已关闭智能检测,所有人声归入对白")
|
| 437 |
-
|
| 438 |
-
algorithm_used = "关闭检测"
|
| 439 |
|
| 440 |
# 4. 分离对白和唱歌
|
|
|
|
| 441 |
status_messages.append("✂️ 正在分离对白和背景音乐...")
|
| 442 |
yield None, None, None, "\n".join(status_messages)
|
| 443 |
|
| 444 |
-
|
| 445 |
|
| 446 |
-
dialog_vocals = vocals *
|
| 447 |
singing_vocals = vocals * singing_mask
|
| 448 |
|
| 449 |
# 5. 生成最终输出
|
|
@@ -462,20 +324,13 @@ def process_audio_full(audio_file, detection_mode, enable_detection):
|
|
| 462 |
output_b = np.clip(instrumental + singing_vocals * singing_gain, -1.0, 1.0)
|
| 463 |
output_c = instrumental
|
| 464 |
|
| 465 |
-
# 保存文件
|
| 466 |
status_messages.append("💾 正在保存输出文件...")
|
| 467 |
yield None, None, None, "\n".join(status_messages)
|
| 468 |
|
| 469 |
-
|
| 470 |
-
|
| 471 |
-
|
| 472 |
-
algo_tag = "Traditional"
|
| 473 |
-
else:
|
| 474 |
-
algo_tag = "NoDetect"
|
| 475 |
-
|
| 476 |
-
path_a = os.path.join(tmpdir, f"A_dialog_{algo_tag}.wav")
|
| 477 |
-
path_b = os.path.join(tmpdir, f"B_bgm_with_singing_{algo_tag}.wav")
|
| 478 |
-
path_c = os.path.join(tmpdir, f"C_instrumental_{algo_tag}.wav")
|
| 479 |
|
| 480 |
save_audio(path_a, output_a, sr)
|
| 481 |
save_audio(path_b, output_b, sr)
|
|
@@ -483,7 +338,7 @@ def process_audio_full(audio_file, detection_mode, enable_detection):
|
|
| 483 |
|
| 484 |
# 统计信息
|
| 485 |
total_duration = len(vocals) / sr
|
| 486 |
-
dialog_duration = np.sum(
|
| 487 |
singing_duration = total_duration - dialog_duration
|
| 488 |
|
| 489 |
status_messages.append("")
|
|
@@ -497,29 +352,11 @@ def process_audio_full(audio_file, detection_mode, enable_detection):
|
|
| 497 |
status_messages.append(f" 音乐人声时长: {singing_duration:.1f} 秒 ({singing_duration/total_duration*100:.1f}%)")
|
| 498 |
status_messages.append(f" 运行设备: {DEVICE.upper()}")
|
| 499 |
status_messages.append("")
|
| 500 |
-
|
| 501 |
-
|
| 502 |
-
|
| 503 |
-
status_messages.append("🎯 本次使用的检测算法:")
|
| 504 |
-
status_messages.append(" ✅✅✅ Silero VAD 深度学习模型")
|
| 505 |
-
status_messages.append(" 📈 准确率: 约 85-90%")
|
| 506 |
-
status_messages.append(" 🧠 技术: 神经网络(10000+ 小时训练)")
|
| 507 |
-
elif algorithm_used == "传统算法":
|
| 508 |
-
status_messages.append("🎯 本次使用的检测算法:")
|
| 509 |
-
status_messages.append(" ⚠️⚠️⚠️ 传统多特征算法(已降级)")
|
| 510 |
-
status_messages.append(" 📉 准确率: 约 75-80%")
|
| 511 |
-
status_messages.append(" 🔧 技术: 能量+零交叉率+频谱+音高")
|
| 512 |
-
status_messages.append("")
|
| 513 |
-
status_messages.append(" ⚠️ 注意: 准确率低于 Silero VAD 约 10-15%")
|
| 514 |
-
status_messages.append(" 💡 如需更高准确率,建议刷新页面重试")
|
| 515 |
-
else:
|
| 516 |
-
status_messages.append("🎯 本次使用的检测算法:")
|
| 517 |
-
status_messages.append(" ⚪ 未启用检测(所有人声归入对白)")
|
| 518 |
-
|
| 519 |
status_messages.append("")
|
| 520 |
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
| 521 |
-
status_messages.append(f"💾 输出文件已标注算法: {algo_tag}")
|
| 522 |
-
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
| 523 |
|
| 524 |
yield (
|
| 525 |
path_a,
|
|
@@ -539,7 +376,7 @@ def process_audio_full(audio_file, detection_mode, enable_detection):
|
|
| 539 |
# 创建 Gradio 界面
|
| 540 |
with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
| 541 |
gr.Markdown(f"""
|
| 542 |
-
# 🎵 AI 音频分离工具 -
|
| 543 |
|
| 544 |
**当前运行设备**: {DEVICE.upper()} {'✅ GPU加速' if DEVICE == 'cuda' else '⚠️ CPU模式'}
|
| 545 |
|
|
@@ -550,38 +387,8 @@ with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
|
| 550 |
|
| 551 |
💡 **核心技术**:
|
| 552 |
- Demucs 4.0 深度学习模型(人声/伴奏分离)
|
| 553 |
-
-
|
| 554 |
-
-
|
| 555 |
-
|
| 556 |
-
---
|
| 557 |
-
|
| 558 |
-
## ⚠️ 重要说明:超时和降级机制
|
| 559 |
-
|
| 560 |
-
### 🔹 什么是"超时"?
|
| 561 |
-
- Silero VAD 模型需要从网络下载(约10MB)
|
| 562 |
-
- 给下载过程设置 **3 分钟(180秒)时间限制**
|
| 563 |
-
- 如果超过 3 分钟还没下载完,就**放弃下载**
|
| 564 |
-
|
| 565 |
-
### 🔹 什么是"降级"?
|
| 566 |
-
- 如果 Silero VAD 下载失败(超时或网络错误)
|
| 567 |
-
- 自动切换到**传统多特征算法**
|
| 568 |
-
- 准确率从 **85-90% 降到 75-80%**
|
| 569 |
-
|
| 570 |
-
### 🔹 如何知道用的是哪个算法?
|
| 571 |
-
1. **处理状态框**会有醒目标注:
|
| 572 |
-
- ✅✅✅ = 使用 Silero VAD(高准确率)
|
| 573 |
-
- ⚠️⚠️⚠️ = 使用传统算法(降级,准确率较低)
|
| 574 |
-
|
| 575 |
-
2. **输出文件名**会包含算法标识:
|
| 576 |
-
- `A_dialog_SileroVAD.wav` = Silero VAD 处理
|
| 577 |
-
- `A_dialog_Traditional.wav` = 传统算法处理
|
| 578 |
-
|
| 579 |
-
3. **最终结果**会明确显示使用的算法和准确率
|
| 580 |
-
|
| 581 |
-
### 💡 如果看到"降级"怎么办?
|
| 582 |
-
- 表示准确率**只有 75-80%**(不是 85-90%)
|
| 583 |
-
- 建议:刷新页面重新尝试
|
| 584 |
-
- 或者:使用稳定版(移除 Silero,准确率稳定在 75-80%)
|
| 585 |
""")
|
| 586 |
|
| 587 |
with gr.Row():
|
|
@@ -603,71 +410,77 @@ with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
|
| 603 |
value=True,
|
| 604 |
label="🎯 启用智能说话检测(推荐开启)"
|
| 605 |
)
|
| 606 |
-
|
| 607 |
-
|
| 608 |
-
|
| 609 |
-
("平衡模式 - 包含部分 Rap/快语", "balanced")
|
| 610 |
-
],
|
| 611 |
-
value="strict",
|
| 612 |
-
label="检测模式"
|
| 613 |
)
|
| 614 |
gr.Markdown("""
|
| 615 |
-
**
|
| 616 |
-
- **
|
| 617 |
-
- **平衡
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 618 |
""")
|
| 619 |
|
| 620 |
-
process_btn = gr.Button("🚀 开始
|
| 621 |
|
| 622 |
with gr.Column(scale=1):
|
| 623 |
status_box = gr.Textbox(
|
| 624 |
-
label="📊 处理状态
|
| 625 |
-
lines=
|
| 626 |
-
max_lines=
|
| 627 |
show_label=True
|
| 628 |
)
|
| 629 |
|
| 630 |
gr.Markdown("---")
|
| 631 |
-
gr.Markdown("## 📥 分离结果
|
| 632 |
|
| 633 |
with gr.Row():
|
| 634 |
-
output_a = gr.Audio(label="🎤 A - 纯对白", type="filepath")
|
| 635 |
-
output_b = gr.Audio(label="🎵 B - 背景音乐+人声", type="filepath")
|
| 636 |
output_c = gr.Audio(label="🎹 C - 纯伴奏", type="filepath")
|
| 637 |
|
| 638 |
process_btn.click(
|
| 639 |
fn=process_audio_full,
|
| 640 |
-
inputs=[audio_input,
|
| 641 |
outputs=[output_a, output_b, output_c, status_box]
|
| 642 |
)
|
| 643 |
|
| 644 |
gr.Markdown("""
|
| 645 |
---
|
| 646 |
-
## 📌
|
| 647 |
|
| 648 |
-
|
| 649 |
-
|---------|--------|------|------|------|
|
| 650 |
-
| **Silero VAD** | **85-90%** | 深度学习,专门训练 | 需要下载(3分钟超时) | ✅✅✅ / SileroVAD |
|
| 651 |
-
| **传统算法** | **75-80%** | 快速稳定,无需下载 | 准确率较低 | ⚠️⚠️⚠️ / Traditional |
|
| 652 |
|
| 653 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 654 |
|
| 655 |
-
##
|
| 656 |
|
| 657 |
-
|
| 658 |
-
-
|
| 659 |
-
-
|
| 660 |
-
-
|
| 661 |
|
| 662 |
-
###
|
| 663 |
-
- 说明 HuggingFace Spaces 网络限制了 Silero 下载
|
| 664 |
-
- 建议使用"稳定版"(移除 Silero,准确率稳定在 75-80%)
|
| 665 |
-
- 或者本地部署(可以手动下载 Silero 模型)
|
| 666 |
|
| 667 |
-
|
| 668 |
-
-
|
| 669 |
-
-
|
| 670 |
-
-
|
| 671 |
""")
|
| 672 |
|
| 673 |
if __name__ == "__main__":
|
|
|
|
| 3 |
import numpy as np
|
| 4 |
import soundfile as sf
|
| 5 |
import librosa
|
|
|
|
| 6 |
|
| 7 |
# 检查 GPU
|
| 8 |
+
try:
|
| 9 |
+
import torch
|
| 10 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 11 |
+
except:
|
| 12 |
+
DEVICE = "cpu"
|
| 13 |
|
| 14 |
+
SAMPLE_RATE = 44100
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
def extract_audio_from_video(video_path, output_path):
|
| 17 |
"""从视频中提取音频"""
|
|
|
|
| 97 |
raise RuntimeError(f"Demucs 分离失败: {str(e)}")
|
| 98 |
|
| 99 |
|
| 100 |
+
def detect_speaking_improved(vocals_audio, sr, strictness=0.6):
|
| 101 |
"""
|
| 102 |
+
改进的说话检测算法(无需外部模型)
|
| 103 |
|
| 104 |
+
基于多特征融合:
|
| 105 |
+
1. 能量包络(RMS)
|
| 106 |
+
2. 零交叉率(ZCR)
|
| 107 |
+
3. 频谱质心(Spectral Centroid)
|
| 108 |
+
4. 频谱滚降(Spectral Rolloff)
|
| 109 |
+
5. 音高连续性
|
| 110 |
|
| 111 |
+
strictness: 0-1,越高越严格(只保留明确的说话)
|
|
|
|
|
|
|
|
|
|
| 112 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 113 |
try:
|
| 114 |
hop_length = 512
|
| 115 |
frame_length = 2048
|
| 116 |
|
| 117 |
+
# ===== 特征1: 能量 =====
|
| 118 |
rms = librosa.feature.rms(y=vocals_audio, frame_length=frame_length, hop_length=hop_length)[0]
|
| 119 |
|
| 120 |
+
# ===== 特征2: 零交叉率 =====
|
| 121 |
zcr = librosa.feature.zero_crossing_rate(vocals_audio, frame_length=frame_length, hop_length=hop_length)[0]
|
| 122 |
|
| 123 |
+
# ===== 特征3: 频谱质心 =====
|
| 124 |
spectral_centroids = librosa.feature.spectral_centroid(y=vocals_audio, sr=sr, hop_length=hop_length)[0]
|
| 125 |
|
| 126 |
+
# ===== 特征4: 频谱滚降 =====
|
| 127 |
+
spectral_rolloff = librosa.feature.spectral_rolloff(y=vocals_audio, sr=sr, hop_length=hop_length)[0]
|
| 128 |
+
|
| 129 |
+
# ===== 特征5: 音高检测 =====
|
| 130 |
try:
|
| 131 |
f0, voiced_flag, voiced_probs = librosa.pyin(
|
| 132 |
vocals_audio,
|
|
|
|
| 137 |
hop_length=hop_length
|
| 138 |
)
|
| 139 |
f0 = np.nan_to_num(f0, nan=0.0)
|
| 140 |
+
voiced_probs = np.nan_to_num(voiced_probs, nan=0.0)
|
| 141 |
except:
|
| 142 |
f0 = np.zeros(len(rms))
|
| 143 |
+
voiced_probs = np.zeros(len(rms))
|
| 144 |
+
|
| 145 |
+
# ===== 特征融合 =====
|
| 146 |
+
min_len = min(len(rms), len(zcr), len(spectral_centroids), len(spectral_rolloff), len(voiced_probs))
|
| 147 |
|
|
|
|
|
|
|
| 148 |
rms = rms[:min_len]
|
| 149 |
zcr = zcr[:min_len]
|
| 150 |
spectral_centroids = spectral_centroids[:min_len]
|
| 151 |
+
spectral_rolloff = spectral_rolloff[:min_len]
|
| 152 |
+
voiced_probs = voiced_probs[:min_len]
|
| 153 |
f0 = f0[:min_len]
|
| 154 |
|
| 155 |
# 说话特征得分
|
| 156 |
+
# 1. 零交叉率高(但不是极高)
|
| 157 |
zcr_score = np.clip((zcr - 0.05) / 0.15, 0, 1)
|
| 158 |
|
| 159 |
+
# 2. 能量适中(不是持续的高能量)
|
| 160 |
rms_norm = rms / (np.max(rms) + 1e-8)
|
| 161 |
energy_variation = np.abs(np.gradient(rms_norm))
|
| 162 |
energy_score = np.clip(energy_variation * 10, 0, 1)
|
| 163 |
|
| 164 |
+
# 3. 频谱质心变化大
|
| 165 |
centroid_variation = np.abs(np.gradient(spectral_centroids))
|
| 166 |
centroid_score = np.clip(centroid_variation / (np.mean(centroid_variation) + 1e-8), 0, 1)
|
| 167 |
|
| 168 |
+
# 4. 音高不连续
|
| 169 |
pitch_continuity = np.zeros_like(f0)
|
| 170 |
for i in range(1, len(f0)):
|
| 171 |
if f0[i] > 0 and f0[i-1] > 0:
|
|
|
|
| 181 |
0.20 * pitch_continuity
|
| 182 |
)
|
| 183 |
|
| 184 |
+
# 根据严格度调整阈值
|
| 185 |
+
threshold = strictness
|
| 186 |
+
speaking_mask = (speaking_score > threshold).astype(np.float32)
|
| 187 |
|
| 188 |
+
# ===== 后处理 =====
|
| 189 |
+
# 去除过短片段(<0.2秒)
|
| 190 |
min_duration = int(0.2 * sr / hop_length)
|
| 191 |
i = 0
|
| 192 |
while i < len(speaking_mask):
|
|
|
|
| 200 |
else:
|
| 201 |
i += 1
|
| 202 |
|
| 203 |
+
# 填充小间隙(<0.15秒)
|
| 204 |
+
gap_threshold = int(0.15 * sr / hop_length)
|
| 205 |
+
i = 0
|
| 206 |
+
while i < len(speaking_mask) - 1:
|
| 207 |
+
if speaking_mask[i] == 1:
|
| 208 |
+
j = i + 1
|
| 209 |
+
while j < len(speaking_mask) and speaking_mask[j] == 0:
|
| 210 |
+
j += 1
|
| 211 |
+
if j < len(speaking_mask) and j - i < gap_threshold:
|
| 212 |
+
speaking_mask[i:j] = 1
|
| 213 |
+
i = j
|
| 214 |
+
else:
|
| 215 |
+
i += 1
|
| 216 |
+
|
| 217 |
+
# 转换为样本级掩码
|
| 218 |
speaking_mask_samples = np.repeat(speaking_mask, hop_length)
|
| 219 |
|
| 220 |
+
# 调整长度
|
| 221 |
if len(speaking_mask_samples) < len(vocals_audio):
|
| 222 |
speaking_mask_samples = np.pad(speaking_mask_samples, (0, len(vocals_audio) - len(speaking_mask_samples)))
|
| 223 |
else:
|
| 224 |
speaking_mask_samples = speaking_mask_samples[:len(vocals_audio)]
|
| 225 |
|
| 226 |
+
# 平滑边界
|
| 227 |
smooth_window = int(0.03 * sr)
|
| 228 |
if smooth_window > 1:
|
| 229 |
speaking_mask_samples = np.convolve(
|
|
|
|
| 236 |
return speaking_mask_samples
|
| 237 |
|
| 238 |
except Exception as e:
|
| 239 |
+
print(f"说话检测失败: {str(e)}")
|
| 240 |
+
import traceback
|
| 241 |
+
traceback.print_exc()
|
| 242 |
+
# 🔴 修复:如果失败,返回全1(假设全是说话),而不是全0
|
| 243 |
+
return np.ones(len(vocals_audio), dtype=np.float32)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 244 |
|
| 245 |
|
| 246 |
+
def process_audio_full(audio_file, strictness, enable_detection):
|
| 247 |
"""完整的音频分离流程"""
|
| 248 |
if audio_file is None:
|
| 249 |
return None, None, None, "❌ 请先上传音频或视频文件"
|
|
|
|
| 269 |
save_audio(temp_wav, audio, sr)
|
| 270 |
|
| 271 |
# 2. Demucs 分离
|
| 272 |
+
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
| 273 |
status_messages.append("🎵 使用 Demucs AI 模型分离人声和伴奏...")
|
| 274 |
status_messages.append(" (首次运行会下载模型,约500MB)")
|
| 275 |
yield None, None, None, "\n".join(status_messages)
|
|
|
|
| 279 |
vocals, _ = librosa.load(vocals_path, sr=sr, mono=True)
|
| 280 |
instrumental, _ = librosa.load(instrumental_path, sr=sr, mono=True)
|
| 281 |
|
| 282 |
+
status_messages.append(" ✅ Demucs 分离完成")
|
| 283 |
+
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
| 284 |
|
| 285 |
+
# 3. 说话检测
|
| 286 |
if enable_detection:
|
| 287 |
+
status_messages.append("")
|
| 288 |
+
status_messages.append("🎤 正在检测说话片段...")
|
| 289 |
+
status_messages.append(" 算法: 多特征融合(能量+零交叉率+频谱+音高)")
|
| 290 |
+
status_messages.append(f" 严格度: {strictness:.2f}")
|
|
|
|
| 291 |
yield None, None, None, "\n".join(status_messages)
|
| 292 |
|
| 293 |
+
# speaking_mask: 1=说话, 0=其他
|
| 294 |
+
speaking_mask = detect_speaking_improved(vocals, sr, strictness)
|
|
|
|
|
|
|
| 295 |
|
| 296 |
+
status_messages.append(" ✅ 检测完成")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 297 |
else:
|
| 298 |
status_messages.append("⚠️ 已关闭智能检测,所有人声归入对白")
|
| 299 |
+
speaking_mask = np.ones(len(vocals), dtype=np.float32)
|
|
|
|
| 300 |
|
| 301 |
# 4. 分离对白和唱歌
|
| 302 |
+
status_messages.append("")
|
| 303 |
status_messages.append("✂️ 正在分离对白和背景音乐...")
|
| 304 |
yield None, None, None, "\n".join(status_messages)
|
| 305 |
|
| 306 |
+
singing_mask = 1 - speaking_mask
|
| 307 |
|
| 308 |
+
dialog_vocals = vocals * speaking_mask
|
| 309 |
singing_vocals = vocals * singing_mask
|
| 310 |
|
| 311 |
# 5. 生成最终输出
|
|
|
|
| 324 |
output_b = np.clip(instrumental + singing_vocals * singing_gain, -1.0, 1.0)
|
| 325 |
output_c = instrumental
|
| 326 |
|
| 327 |
+
# 保存文件
|
| 328 |
status_messages.append("💾 正在保存输出文件...")
|
| 329 |
yield None, None, None, "\n".join(status_messages)
|
| 330 |
|
| 331 |
+
path_a = os.path.join(tmpdir, "A_dialog.wav")
|
| 332 |
+
path_b = os.path.join(tmpdir, "B_bgm_with_singing.wav")
|
| 333 |
+
path_c = os.path.join(tmpdir, "C_instrumental.wav")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 334 |
|
| 335 |
save_audio(path_a, output_a, sr)
|
| 336 |
save_audio(path_b, output_b, sr)
|
|
|
|
| 338 |
|
| 339 |
# 统计信息
|
| 340 |
total_duration = len(vocals) / sr
|
| 341 |
+
dialog_duration = np.sum(speaking_mask) / sr
|
| 342 |
singing_duration = total_duration - dialog_duration
|
| 343 |
|
| 344 |
status_messages.append("")
|
|
|
|
| 352 |
status_messages.append(f" 音乐人声时长: {singing_duration:.1f} 秒 ({singing_duration/total_duration*100:.1f}%)")
|
| 353 |
status_messages.append(f" 运行设备: {DEVICE.upper()}")
|
| 354 |
status_messages.append("")
|
| 355 |
+
status_messages.append("🎯 检测算法: 传统多特征融合")
|
| 356 |
+
status_messages.append(" 📈 预期准确率: 75-80%")
|
| 357 |
+
status_messages.append(" 🔧 技术: 能量+零交叉率+频谱+音高")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 358 |
status_messages.append("")
|
| 359 |
status_messages.append("━━━━━━━━━━━━━━━━━━━━")
|
|
|
|
|
|
|
| 360 |
|
| 361 |
yield (
|
| 362 |
path_a,
|
|
|
|
| 376 |
# 创建 Gradio 界面
|
| 377 |
with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
| 378 |
gr.Markdown(f"""
|
| 379 |
+
# 🎵 AI 音频分离工具 - 稳定版
|
| 380 |
|
| 381 |
**当前运行设备**: {DEVICE.upper()} {'✅ GPU加速' if DEVICE == 'cuda' else '⚠️ CPU模式'}
|
| 382 |
|
|
|
|
| 387 |
|
| 388 |
💡 **核心技术**:
|
| 389 |
- Demucs 4.0 深度学习模型(人声/伴奏分离)
|
| 390 |
+
- 多特征融合算法(能量、零交叉率、频谱、音高)
|
| 391 |
+
- **准确率 75-80%,稳定快速**
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 392 |
""")
|
| 393 |
|
| 394 |
with gr.Row():
|
|
|
|
| 410 |
value=True,
|
| 411 |
label="🎯 启用智能说话检测(推荐开启)"
|
| 412 |
)
|
| 413 |
+
strictness = gr.Slider(
|
| 414 |
+
0.4, 0.8, value=0.6, step=0.05,
|
| 415 |
+
label="检测严格度"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 416 |
)
|
| 417 |
gr.Markdown("""
|
| 418 |
+
**调节建议**:
|
| 419 |
+
- **0.45-0.55**: 宽松(更多人声归入对白)
|
| 420 |
+
- **0.60-0.65**: 平衡(**推荐**,默认0.60)
|
| 421 |
+
- **0.70-0.80**: 严格(只保留明确的说话)
|
| 422 |
+
|
| 423 |
+
**效果不满意?试试这样调**:
|
| 424 |
+
- 说话被误判为唱歌 → 降低到 0.50-0.55
|
| 425 |
+
- 唱歌被误判为说话 → 提高到 0.70-0.75
|
| 426 |
""")
|
| 427 |
|
| 428 |
+
process_btn = gr.Button("🚀 开始智能分离", variant="primary", size="lg")
|
| 429 |
|
| 430 |
with gr.Column(scale=1):
|
| 431 |
status_box = gr.Textbox(
|
| 432 |
+
label="📊 处理状态",
|
| 433 |
+
lines=20,
|
| 434 |
+
max_lines=25,
|
| 435 |
show_label=True
|
| 436 |
)
|
| 437 |
|
| 438 |
gr.Markdown("---")
|
| 439 |
+
gr.Markdown("## 📥 分离结果")
|
| 440 |
|
| 441 |
with gr.Row():
|
| 442 |
+
output_a = gr.Audio(label="🎤 A - 纯对白(旁白/解说)", type="filepath")
|
| 443 |
+
output_b = gr.Audio(label="🎵 B - 背景音乐+人声(含唱歌/Rap)", type="filepath")
|
| 444 |
output_c = gr.Audio(label="🎹 C - 纯伴奏", type="filepath")
|
| 445 |
|
| 446 |
process_btn.click(
|
| 447 |
fn=process_audio_full,
|
| 448 |
+
inputs=[audio_input, strictness, enable_detection],
|
| 449 |
outputs=[output_a, output_b, output_c, status_box]
|
| 450 |
)
|
| 451 |
|
| 452 |
gr.Markdown("""
|
| 453 |
---
|
| 454 |
+
## 📌 使用说明
|
| 455 |
|
| 456 |
+
### 🎯 本版本特点
|
|
|
|
|
|
|
|
|
|
| 457 |
|
| 458 |
+
- ✅ **稳定快速**:无需下载外部模型
|
| 459 |
+
- ✅ **准确率 75-80%**:适合大部分场景
|
| 460 |
+
- ✅ **修复BUG**:确保对白始终有人声
|
| 461 |
+
- ✅ **启动快速**:3-5分钟构建完成
|
| 462 |
+
|
| 463 |
+
### 💡 如何获得最佳效果
|
| 464 |
+
|
| 465 |
+
1. **优先用默认值 0.60** 测试
|
| 466 |
+
2. 根据结果微调严格度:
|
| 467 |
+
- 对白太少 → 降低到 0.50-0.55
|
| 468 |
+
- 对白太多 → 提高到 0.70-0.75
|
| 469 |
+
3. 每次调整 0.05 观察变化
|
| 470 |
|
| 471 |
+
### ⚠️ 技术限制
|
| 472 |
|
| 473 |
+
传统算法准确率有限,以下情况仍有挑战:
|
| 474 |
+
- 说唱风格旁白
|
| 475 |
+
- 快速说话 + 背景音乐
|
| 476 |
+
- 唱歌式说话
|
| 477 |
|
| 478 |
+
### 🔬 如果需要更高准确率
|
|
|
|
|
|
|
|
|
|
| 479 |
|
| 480 |
+
可以考虑:
|
| 481 |
+
- 使用专业软件(如 Adobe Audition)
|
| 482 |
+
- 本地部署并手动下载 Silero VAD 模型
|
| 483 |
+
- 训练深度学习分类模型
|
| 484 |
""")
|
| 485 |
|
| 486 |
if __name__ == "__main__":
|