add custom vad silence

transcribe/pipelines/pipe_vad.py

@@ -2,10 +2,27 @@
 from .base import MetaItem, BasePipe
 from ..helpers.vadprocessor import SileroVADProcessor, FixedVADIterator
 import numpy as np
-from silero_vad import get_speech_timestamps,collect_chunks
+from silero_vad import get_speech_timestamps
 import torch
+from typing import List
 # import noisereduce as nr
 
+def collect_chunks(tss: List[dict], wav: torch.Tensor, sample_rate: int = 16000):
+    chunks = []
+    silent_samples = int(0.3 * sample_rate)  # 300ms 的静音样本数
+    silence = torch.zeros(silent_samples)  # 创建300ms的静音
+    
+    for i in range(len(tss)):
+        # 先添加当前语音片段
+        chunks.append(wav[tss[i]['start']: tss[i]['end']])
+        
+        # 如果不是最后一个片段，且与下一个片段间隔大于100ms，则添加静音
+        if i < len(tss) - 1:
+            gap = tss[i+1]['start'] - tss[i]['end']
+            if gap > 0.1 * sample_rate:  # 判断间隔是否大于100ms
+                chunks.append(silence)  # 添加300ms静音
+    
+    return torch.cat(chunks)
 
 class VadPipe(BasePipe):
     model = None