wx44wx/py-active-call
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add energy based vad fallback

Browse Source
This commit is contained in:
Xin Wang
2026-02-05 17:21:52 +08:00
parent b72e09f263
commit 2de427b92c
2 changed files with 14 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

BIN
data/response.wav Normal file
View File

Binary file not shown.

16
processors/vad.py
View File

@@ -82,8 +82,20 @@ class SileroVAD:
Tuple of (label, probability) where label is "Speech" or "Silence" Tuple of (label, probability) where label is "Speech" or "Silence"
""" """
if self.session is None or not ONNX_AVAILABLE: if self.session is None or not ONNX_AVAILABLE:
# If model not loaded or onnxruntime not available, assume speech # Fallback energy-based VAD when model isn't available.
return "Speech", 1.0 # Map RMS energy to a pseudo-probability so the existing threshold works.
if not pcm_bytes:
return "Silence", 0.0
audio_int16 = np.frombuffer(pcm_bytes, dtype=np.int16)
if audio_int16.size == 0:
return "Silence", 0.0
audio_float = audio_int16.astype(np.float32) / 32768.0
rms = float(np.sqrt(np.mean(audio_float * audio_float)))
# Typical speech RMS is ~0.02-0.05 at 16-bit normalized scale.
# Normalize so threshold=0.5 roughly corresponds to ~0.025 RMS.
probability = min(1.0, rms / 0.05)
label = "Speech" if probability >= 0.5 else "Silence"
return label, probability
# Convert bytes to numpy array of int16 # Convert bytes to numpy array of int16
audio_int16 = np.frombuffer(pcm_bytes, dtype=np.int16) audio_int16 = np.frombuffer(pcm_bytes, dtype=np.int16)