Spaces:
Sleeping
Sleeping
Upload 3 files
Browse files- app.py +116 -121
- requirements.txt +1 -2
app.py
CHANGED
|
@@ -5,18 +5,34 @@ import soundfile as sf
|
|
| 5 |
import librosa
|
| 6 |
import torch
|
| 7 |
|
| 8 |
-
# 加载 Silero VAD 模型(用于检测说话)
|
| 9 |
-
try:
|
| 10 |
-
from silero_vad import load_silero_vad, get_speech_timestamps
|
| 11 |
-
SILERO_AVAILABLE = True
|
| 12 |
-
except:
|
| 13 |
-
SILERO_AVAILABLE = False
|
| 14 |
-
print("⚠️ Silero VAD 不可用,将使用传统算法")
|
| 15 |
-
|
| 16 |
# 检查 GPU
|
| 17 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 18 |
SAMPLE_RATE = 44100
|
| 19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
def extract_audio_from_video(video_path, output_path):
|
| 21 |
"""从视频中提取音频"""
|
| 22 |
try:
|
|
@@ -107,6 +123,10 @@ def detect_speech_with_silero(vocals_audio, sr):
|
|
| 107 |
返回:speech_mask (1=说话, 0=其他)
|
| 108 |
"""
|
| 109 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
# 重采样到 16kHz(Silero VAD 要求)
|
| 111 |
if sr != 16000:
|
| 112 |
vocals_16k = librosa.resample(vocals_audio, orig_sr=sr, target_sr=16000)
|
|
@@ -115,31 +135,25 @@ def detect_speech_with_silero(vocals_audio, sr):
|
|
| 115 |
vocals_16k = vocals_audio
|
| 116 |
sr_work = 16000
|
| 117 |
|
| 118 |
-
# 加载模型
|
| 119 |
-
model = load_silero_vad()
|
| 120 |
-
|
| 121 |
# 转换为 torch tensor
|
| 122 |
-
audio_tensor = torch.from_numpy(vocals_16k).float()
|
| 123 |
|
| 124 |
-
#
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
model,
|
| 128 |
-
threshold=0.5, # 检测阈值
|
| 129 |
-
sampling_rate=sr_work,
|
| 130 |
-
min_speech_duration_ms=250, # 最短说话时长
|
| 131 |
-
min_silence_duration_ms=100, # 最短静音时长
|
| 132 |
-
window_size_samples=512,
|
| 133 |
-
speech_pad_ms=30
|
| 134 |
-
)
|
| 135 |
|
| 136 |
-
|
| 137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
speech_mask[start:end] = 1.0
|
| 143 |
|
| 144 |
# 调整回原始采样率
|
| 145 |
if sr != sr_work:
|
|
@@ -164,50 +178,12 @@ def detect_speech_with_silero(vocals_audio, sr):
|
|
| 164 |
print(f"Silero VAD 检测失败: {str(e)}")
|
| 165 |
import traceback
|
| 166 |
traceback.print_exc()
|
| 167 |
-
# 失败时返回全零
|
| 168 |
return np.zeros(len(vocals_audio), dtype=np.float32)
|
| 169 |
|
| 170 |
|
| 171 |
-
def detect_singing_hybrid(vocals_audio, sr, mode='strict'):
|
| 172 |
-
"""
|
| 173 |
-
混合检测策略:
|
| 174 |
-
1. 先用 Silero VAD 检测"说话"
|
| 175 |
-
2. 其余全部归入"唱歌/音乐"
|
| 176 |
-
|
| 177 |
-
mode='strict': 严格模式,只有明确的说话才归入对白
|
| 178 |
-
mode='balanced': 平衡模式,包含部分 Rap
|
| 179 |
-
"""
|
| 180 |
-
try:
|
| 181 |
-
if SILERO_AVAILABLE:
|
| 182 |
-
print("🎯 使用 Silero VAD 深度学习模型检测说话...")
|
| 183 |
-
speech_mask = detect_speech_with_silero(vocals_audio, sr)
|
| 184 |
-
else:
|
| 185 |
-
print("⚠️ Silero 不可用,使用传统算法...")
|
| 186 |
-
speech_mask = detect_speech_fallback(vocals_audio, sr)
|
| 187 |
-
|
| 188 |
-
if mode == 'strict':
|
| 189 |
-
# 严格模式:只保留明确的说话
|
| 190 |
-
# 缩小说话区域,避免误判
|
| 191 |
-
from scipy.ndimage import binary_erosion
|
| 192 |
-
kernel_size = int(0.05 * sr) # 50ms
|
| 193 |
-
if kernel_size > 1:
|
| 194 |
-
speech_mask = binary_erosion(speech_mask, structure=np.ones(kernel_size)).astype(np.float32)
|
| 195 |
-
|
| 196 |
-
# 说话 = 1, 唱歌 = 0
|
| 197 |
-
# 我们需要返回唱歌掩码,所以要反转
|
| 198 |
-
singing_mask = 1 - speech_mask
|
| 199 |
-
|
| 200 |
-
return singing_mask
|
| 201 |
-
|
| 202 |
-
except Exception as e:
|
| 203 |
-
print(f"检测失败: {str(e)}")
|
| 204 |
-
return np.ones(len(vocals_audio), dtype=np.float32) # 全部归入唱歌
|
| 205 |
-
|
| 206 |
-
|
| 207 |
def detect_speech_fallback(vocals_audio, sr):
|
| 208 |
"""传统算法备用方案(当 Silero 不可用时)"""
|
| 209 |
try:
|
| 210 |
-
# 使用能量 + 零交叉率检测说话
|
| 211 |
hop_length = 512
|
| 212 |
|
| 213 |
# 能量
|
|
@@ -234,6 +210,40 @@ def detect_speech_fallback(vocals_audio, sr):
|
|
| 234 |
return np.zeros(len(vocals_audio), dtype=np.float32)
|
| 235 |
|
| 236 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 237 |
def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
| 238 |
"""完整的音频分离流程"""
|
| 239 |
if audio_file is None:
|
|
@@ -243,6 +253,17 @@ def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
|
| 243 |
|
| 244 |
try:
|
| 245 |
with tempfile.TemporaryDirectory() as tmpdir:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 246 |
# 1. 加载音频
|
| 247 |
status_messages.append("📂 正在加载文件...")
|
| 248 |
yield None, None, None, "\n".join(status_messages)
|
|
@@ -261,10 +282,6 @@ def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
|
| 261 |
|
| 262 |
# 2. Demucs 分离
|
| 263 |
status_messages.append("🎵 使用 Demucs AI 模型分离人声和伴奏...")
|
| 264 |
-
if SILERO_AVAILABLE:
|
| 265 |
-
status_messages.append(" ✅ 已启用 Silero VAD 深度学习检测器")
|
| 266 |
-
else:
|
| 267 |
-
status_messages.append(" ⚠️ 使用传统算法(准确率较低)")
|
| 268 |
yield None, None, None, "\n".join(status_messages)
|
| 269 |
|
| 270 |
vocals_path, instrumental_path = run_demucs_separation(temp_wav, tmpdir)
|
|
@@ -274,10 +291,9 @@ def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
|
| 274 |
|
| 275 |
# 3. 说话/唱歌检测
|
| 276 |
if enable_singing_detection:
|
| 277 |
-
status_messages.append("🎤
|
| 278 |
yield None, None, None, "\n".join(status_messages)
|
| 279 |
|
| 280 |
-
# singing_mask: 1=唱歌, 0=说话
|
| 281 |
singing_mask = detect_singing_hybrid(vocals, sr, mode=detection_mode)
|
| 282 |
else:
|
| 283 |
status_messages.append("⚠️ 已关闭智能检测,所有人声归入对白")
|
|
@@ -287,7 +303,7 @@ def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
|
| 287 |
status_messages.append("✂️ 正在分离对白和背景音乐...")
|
| 288 |
yield None, None, None, "\n".join(status_messages)
|
| 289 |
|
| 290 |
-
dialog_mask = 1 - singing_mask
|
| 291 |
|
| 292 |
dialog_vocals = vocals * dialog_mask
|
| 293 |
singing_vocals = vocals * singing_mask
|
|
@@ -334,12 +350,10 @@ def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
|
| 334 |
status_messages.append(f" 运行设备: {DEVICE.upper()}")
|
| 335 |
|
| 336 |
if enable_singing_detection:
|
| 337 |
-
if
|
| 338 |
status_messages.append(f"\n💡 检测算法: Silero VAD 深度学习")
|
| 339 |
-
status_messages.append(f" 准确率: 约 85-90%")
|
| 340 |
else:
|
| 341 |
status_messages.append(f"\n💡 检测算法: 传统信号处理")
|
| 342 |
-
status_messages.append(f" 准确率: 约 70-75%")
|
| 343 |
|
| 344 |
status_messages.append(f"━━━━━━━━━━━━━━━━━━━━")
|
| 345 |
|
|
@@ -363,17 +377,16 @@ with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
|
| 363 |
gr.Markdown(f"""
|
| 364 |
# 🎵 AI 音频分离工具 - 深度学习版
|
| 365 |
|
| 366 |
-
**当前运行设备**: {DEVICE.upper()} {'✅ GPU加速' if DEVICE == 'cuda' else '⚠️ CPU模式'}
|
| 367 |
-
**AI检测器**: {'✅ Silero VAD (深度学习)' if SILERO_AVAILABLE else '⚠️ 传统算法'}
|
| 368 |
|
| 369 |
-
##
|
| 370 |
- **A - 纯对白**: 旁白、解说、对话(不含Rap/口号)
|
| 371 |
- **B - 背景音乐+人声**: 伴奏 + 唱歌 + Rap + 和声
|
| 372 |
- **C - 纯伴奏**: 去除所有人声的纯音乐
|
| 373 |
|
| 374 |
💡 **核心技术**:
|
| 375 |
- Demucs 4.0 深度学习模型(人声/伴奏分离)
|
| 376 |
-
- Silero VAD
|
| 377 |
""")
|
| 378 |
|
| 379 |
with gr.Row():
|
|
@@ -405,7 +418,7 @@ with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
|
| 405 |
)
|
| 406 |
gr.Markdown("""
|
| 407 |
**模式说明**:
|
| 408 |
-
-
|
| 409 |
- **平衡模式**:包含部分 Rap 风格的说话
|
| 410 |
|
| 411 |
💡 **大部分场景用严格模式效果最好!**
|
|
@@ -437,53 +450,35 @@ with gr.Blocks(theme=gr.themes.Soft(), title="AI音频分离工具") as demo:
|
|
| 437 |
|
| 438 |
gr.Markdown("""
|
| 439 |
---
|
| 440 |
-
## 📌
|
| 441 |
-
|
| 442 |
-
### 🎯 为什么改成"纯对白"定义
|
| 443 |
-
|
| 444 |
-
根据实际测试,我们发现:
|
| 445 |
-
- **Rap 介于说话和唱歌之间**,传统算法很难区分
|
| 446 |
-
- **大部分用户真正需要的是"旁白/解说"**,而不是 Rap
|
| 447 |
-
- **唱歌检测的核心难点在于 Rap**(它有节奏但不是旋律)
|
| 448 |
-
|
| 449 |
-
因此新版本:
|
| 450 |
-
- ✅ A区域:只保留纯说话(旁白、对话、解说)
|
| 451 |
-
- ✅ B区域:包含所有"有节奏感的人声"(唱歌、Rap、和声、口号)
|
| 452 |
-
- ✅ C区域:纯音乐(无人声)
|
| 453 |
-
|
| 454 |
-
### 🧠 Silero VAD 深度学习模型
|
| 455 |
|
| 456 |
-
|
| 457 |
-
- **准确率**: 说话检测准确率 85-90%
|
| 458 |
-
- **优势**: 专门训练识别"自然说话",对 Rap/唱歌免疫
|
| 459 |
-
- **开源**: 完全免费,MIT 协议
|
| 460 |
|
| 461 |
-
|
|
|
|
|
|
|
|
|
|
| 462 |
|
| 463 |
-
|
| 464 |
-
|
| 465 |
-
|
| 466 |
-
|
| 467 |
|
| 468 |
-
|
|
|
|
|
|
|
| 469 |
|
| 470 |
-
###
|
| 471 |
|
| 472 |
-
|
| 473 |
-
|
| 474 |
-
|
| 475 |
-
- 在专业音频软件中手动编辑
|
| 476 |
-
- 使用付费商业软件(如 Adobe Audition)
|
| 477 |
-
- 训练专门的分类模型(需要大量数据)
|
| 478 |
|
| 479 |
-
###
|
| 480 |
|
| 481 |
-
|
| 482 |
-
|
| 483 |
-
|
| 484 |
-
| 多特征融合 | 70-75% | 准确率提升 | 仍难处理边缘情况 |
|
| 485 |
-
| **Silero VAD** | **85-90%** | **专���训练** | **需要网络下载模型** |
|
| 486 |
-
| 商业软件 | 95%+ | 接近完美 | 付费、闭源 |
|
| 487 |
""")
|
| 488 |
|
| 489 |
if __name__ == "__main__":
|
|
|
|
| 5 |
import librosa
|
| 6 |
import torch
|
| 7 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
# 检查 GPU
|
| 9 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 10 |
SAMPLE_RATE = 44100
|
| 11 |
|
| 12 |
+
# 全局加载 Silero VAD 模型
|
| 13 |
+
SILERO_MODEL = None
|
| 14 |
+
|
| 15 |
+
def load_silero_model():
|
| 16 |
+
"""加载 Silero VAD 模型"""
|
| 17 |
+
global SILERO_MODEL
|
| 18 |
+
if SILERO_MODEL is None:
|
| 19 |
+
try:
|
| 20 |
+
print("📥 正在下载 Silero VAD 模型...")
|
| 21 |
+
SILERO_MODEL, utils = torch.hub.load(
|
| 22 |
+
repo_or_dir='snakers4/silero-vad',
|
| 23 |
+
model='silero_vad',
|
| 24 |
+
force_reload=False,
|
| 25 |
+
onnx=False
|
| 26 |
+
)
|
| 27 |
+
SILERO_MODEL = SILERO_MODEL.to(DEVICE)
|
| 28 |
+
print("✅ Silero VAD 模型加载成功")
|
| 29 |
+
return True
|
| 30 |
+
except Exception as e:
|
| 31 |
+
print(f"⚠️ Silero VAD 加载失败: {str(e)}")
|
| 32 |
+
SILERO_MODEL = None
|
| 33 |
+
return False
|
| 34 |
+
return True
|
| 35 |
+
|
| 36 |
def extract_audio_from_video(video_path, output_path):
|
| 37 |
"""从视频中提取音频"""
|
| 38 |
try:
|
|
|
|
| 123 |
返回:speech_mask (1=说话, 0=其他)
|
| 124 |
"""
|
| 125 |
try:
|
| 126 |
+
global SILERO_MODEL
|
| 127 |
+
if SILERO_MODEL is None:
|
| 128 |
+
raise RuntimeError("Silero 模型未加载")
|
| 129 |
+
|
| 130 |
# 重采样到 16kHz(Silero VAD 要求)
|
| 131 |
if sr != 16000:
|
| 132 |
vocals_16k = librosa.resample(vocals_audio, orig_sr=sr, target_sr=16000)
|
|
|
|
| 135 |
vocals_16k = vocals_audio
|
| 136 |
sr_work = 16000
|
| 137 |
|
|
|
|
|
|
|
|
|
|
| 138 |
# 转换为 torch tensor
|
| 139 |
+
audio_tensor = torch.from_numpy(vocals_16k).float().to(DEVICE)
|
| 140 |
|
| 141 |
+
# 使用 Silero VAD 检测
|
| 142 |
+
window_size_samples = 512
|
| 143 |
+
speech_probs = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
|
| 145 |
+
for i in range(0, len(audio_tensor), window_size_samples):
|
| 146 |
+
chunk = audio_tensor[i:i+window_size_samples]
|
| 147 |
+
if len(chunk) < window_size_samples:
|
| 148 |
+
chunk = torch.nn.functional.pad(chunk, (0, window_size_samples - len(chunk)))
|
| 149 |
+
|
| 150 |
+
with torch.no_grad():
|
| 151 |
+
speech_prob = SILERO_MODEL(chunk.unsqueeze(0), sr_work).item()
|
| 152 |
+
speech_probs.append(speech_prob)
|
| 153 |
|
| 154 |
+
# 创建掩码
|
| 155 |
+
speech_mask = np.repeat(speech_probs, window_size_samples)[:len(vocals_16k)]
|
| 156 |
+
speech_mask = (speech_mask > 0.5).astype(np.float32)
|
|
|
|
| 157 |
|
| 158 |
# 调整回原始采样率
|
| 159 |
if sr != sr_work:
|
|
|
|
| 178 |
print(f"Silero VAD 检测失败: {str(e)}")
|
| 179 |
import traceback
|
| 180 |
traceback.print_exc()
|
|
|
|
| 181 |
return np.zeros(len(vocals_audio), dtype=np.float32)
|
| 182 |
|
| 183 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 184 |
def detect_speech_fallback(vocals_audio, sr):
|
| 185 |
"""传统算法备用方案(当 Silero 不可用时)"""
|
| 186 |
try:
|
|
|
|
| 187 |
hop_length = 512
|
| 188 |
|
| 189 |
# 能量
|
|
|
|
| 210 |
return np.zeros(len(vocals_audio), dtype=np.float32)
|
| 211 |
|
| 212 |
|
| 213 |
+
def detect_singing_hybrid(vocals_audio, sr, mode='strict'):
|
| 214 |
+
"""
|
| 215 |
+
混合检测策略:
|
| 216 |
+
1. 先用 Silero VAD 检测"说话"
|
| 217 |
+
2. 其余全部归入"唱歌/音乐"
|
| 218 |
+
"""
|
| 219 |
+
try:
|
| 220 |
+
global SILERO_MODEL
|
| 221 |
+
|
| 222 |
+
if SILERO_MODEL is not None:
|
| 223 |
+
print("🎯 使用 Silero VAD 深度学习模型检测说话...")
|
| 224 |
+
speech_mask = detect_speech_with_silero(vocals_audio, sr)
|
| 225 |
+
else:
|
| 226 |
+
print("⚠️ Silero 不可用,使用传统算法...")
|
| 227 |
+
speech_mask = detect_speech_fallback(vocals_audio, sr)
|
| 228 |
+
|
| 229 |
+
if mode == 'strict':
|
| 230 |
+
# 严格模式:只保留明确的说话
|
| 231 |
+
from scipy.ndimage import binary_erosion
|
| 232 |
+
kernel_size = int(0.05 * sr) # 50ms
|
| 233 |
+
if kernel_size > 1:
|
| 234 |
+
speech_mask = binary_erosion(speech_mask, structure=np.ones(kernel_size)).astype(np.float32)
|
| 235 |
+
|
| 236 |
+
# 说话 = 1, 唱歌 = 0
|
| 237 |
+
# 返回唱歌掩码
|
| 238 |
+
singing_mask = 1 - speech_mask
|
| 239 |
+
|
| 240 |
+
return singing_mask
|
| 241 |
+
|
| 242 |
+
except Exception as e:
|
| 243 |
+
print(f"检测失败: {str(e)}")
|
| 244 |
+
return np.ones(len(vocals_audio), dtype=np.float32)
|
| 245 |
+
|
| 246 |
+
|
| 247 |
def process_audio_full(audio_file, detection_mode, enable_singing_detection):
|
| 248 |
"""完整的音频分离流程"""
|
| 249 |
if audio_file is None:
|
|
|
|
| 253 |
|
| 254 |
try:
|
| 255 |
with tempfile.TemporaryDirectory() as tmpdir:
|
| 256 |
+
# 0. 加载 Silero 模型(如果需要)
|
| 257 |
+
if enable_singing_detection:
|
| 258 |
+
status_messages.append("🔧 正在初始化 AI 检测器...")
|
| 259 |
+
yield None, None, None, "\n".join(status_messages)
|
| 260 |
+
silero_loaded = load_silero_model()
|
| 261 |
+
if silero_loaded:
|
| 262 |
+
status_messages.append(" ✅ Silero VAD 深度学习模型已就绪")
|
| 263 |
+
else:
|
| 264 |
+
status_messages.append(" ⚠️ Silero 加载失败,将使用传统算法")
|
| 265 |
+
yield None, None, None, "\n".join(status_messages)
|
| 266 |
+
|
| 267 |
# 1. 加载音频
|
| 268 |
status_messages.append("📂 正在加载文件...")
|
| 269 |
yield None, None, None, "\n".join(status_messages)
|
|
|
|
| 282 |
|
| 283 |
# 2. Demucs 分离
|
| 284 |
status_messages.append("🎵 使用 Demucs AI 模型分离人声和伴奏...")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 285 |
yield None, None, None, "\n".join(status_messages)
|
| 286 |
|
| 287 |
vocals_path, instrumental_path = run_demucs_separation(temp_wav, tmpdir)
|
|
|
|
| 291 |
|
| 292 |
# 3. 说话/唱歌检测
|
| 293 |
if enable_singing_detection:
|
| 294 |
+
status_messages.append("🎤 正在检测说话片段...")
|
| 295 |
yield None, None, None, "\n".join(status_messages)
|
| 296 |
|
|
|
|
| 297 |
singing_mask = detect_singing_hybrid(vocals, sr, mode=detection_mode)
|
| 298 |
else:
|
| 299 |
status_messages.append("⚠️ 已关闭智能检测,所有人声归入对白")
|
|
|
|
| 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 |
+
dialog_mask = 1 - singing_mask
|
| 307 |
|
| 308 |
dialog_vocals = vocals * dialog_mask
|
| 309 |
singing_vocals = vocals * singing_mask
|
|
|
|
| 350 |
status_messages.append(f" 运行设备: {DEVICE.upper()}")
|
| 351 |
|
| 352 |
if enable_singing_detection:
|
| 353 |
+
if SILERO_MODEL is not None:
|
| 354 |
status_messages.append(f"\n💡 检测算法: Silero VAD 深度学习")
|
|
|
|
| 355 |
else:
|
| 356 |
status_messages.append(f"\n💡 检测算法: 传统信号处理")
|
|
|
|
| 357 |
|
| 358 |
status_messages.append(f"━━━━━━━━━━━━━━━━━━━━")
|
| 359 |
|
|
|
|
| 377 |
gr.Markdown(f"""
|
| 378 |
# 🎵 AI 音频分离工具 - 深度学习版
|
| 379 |
|
| 380 |
+
**当前运行设备**: {DEVICE.upper()} {'✅ GPU加速' if DEVICE == 'cuda' else '⚠️ CPU模式'}
|
|
|
|
| 381 |
|
| 382 |
+
## 功能说明
|
| 383 |
- **A - 纯对白**: 旁白、解说、对话(不含Rap/口号)
|
| 384 |
- **B - 背景音乐+人声**: 伴奏 + 唱歌 + Rap + 和声
|
| 385 |
- **C - 纯伴奏**: 去除所有人声的纯音乐
|
| 386 |
|
| 387 |
💡 **核心技术**:
|
| 388 |
- Demucs 4.0 深度学习模型(人声/伴奏分离)
|
| 389 |
+
- Silero VAD 神经网络(说话检测)
|
| 390 |
""")
|
| 391 |
|
| 392 |
with gr.Row():
|
|
|
|
| 418 |
)
|
| 419 |
gr.Markdown("""
|
| 420 |
**模式说明**:
|
| 421 |
+
- **严格模式**(推荐):只有清晰的说话才归入对白
|
| 422 |
- **平衡模式**:包含部分 Rap 风格的说话
|
| 423 |
|
| 424 |
💡 **大部分场景用严格模式效果最好!**
|
|
|
|
| 450 |
|
| 451 |
gr.Markdown("""
|
| 452 |
---
|
| 453 |
+
## 📌 使用说明
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 454 |
|
| 455 |
+
### 🎯 新版本改进
|
|
|
|
|
|
|
|
|
|
| 456 |
|
| 457 |
+
1. **使用 Silero VAD 深度学习模型**
|
| 458 |
+
- 自动从 torch.hub 下载(约10MB)
|
| 459 |
+
- 准确率比传统算法提升 15-20%
|
| 460 |
+
- 专门训练识别"说话"
|
| 461 |
|
| 462 |
+
2. **改变产品定义**
|
| 463 |
+
- A区域:只保留纯说话(旁白、对话)
|
| 464 |
+
- B区域:所有音乐性人声(唱歌、Rap、和声)
|
| 465 |
+
- 逻辑更清晰,用户需求更明确
|
| 466 |
|
| 467 |
+
3. **两种检测模式**
|
| 468 |
+
- 严格模式:优先保证对白纯净度
|
| 469 |
+
- 平衡模式:包含部分快速说话
|
| 470 |
|
| 471 |
+
### ⚠️ 技术限制
|
| 472 |
|
| 473 |
+
- **深度学习准确率**: 85-90%(已是免费方案极限)
|
| 474 |
+
- **边缘情况**: 说唱风格旁白、唱歌式说话仍有挑战
|
| 475 |
+
- **完美分离**: 需要付费商业软件或自训练模型
|
|
|
|
|
|
|
|
|
|
| 476 |
|
| 477 |
+
### 💡 效果不满意?
|
| 478 |
|
| 479 |
+
1. 尝试两种模式切换
|
| 480 |
+
2. 在专业音频软件中手动微调(推荐 Audacity)
|
| 481 |
+
3. 考虑使用付费商业软件(如 Adobe Audition)
|
|
|
|
|
|
|
|
|
|
| 482 |
""")
|
| 483 |
|
| 484 |
if __name__ == "__main__":
|
requirements.txt
CHANGED
|
@@ -5,5 +5,4 @@ torchaudio==2.1.0
|
|
| 5 |
librosa==0.10.1
|
| 6 |
soundfile==0.12.1
|
| 7 |
numpy==1.24.3
|
| 8 |
-
scipy==1.11.4
|
| 9 |
-
silero-vad==4.0.0
|
|
|
|
| 5 |
librosa==0.10.1
|
| 6 |
soundfile==0.12.1
|
| 7 |
numpy==1.24.3
|
| 8 |
+
scipy==1.11.4
|
|
|