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

Fix microphone talk eou missing and clean chat log

Browse Source
This commit is contained in:
Xin Wang
2026-02-06 11:36:39 +08:00
parent c8c0e30bc3
commit 7be8fda424
2 changed files with 21 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
examples/web_client.html
View File

@@ -719,7 +719,14 @@
sendCommand({ command: "chat", text }); sendCommand({ command: "chat", text });
chatInput.value = ""; chatInput.value = "";
}); });
clearLogBtn.addEventListener("click", () => (logEl.innerHTML = "")); clearLogBtn.addEventListener("click", () => {
logEl.innerHTML = "";
chatHistory.innerHTML = "";
setInterim("You", "");
setInterim("AI", "");
interimUserText = "";
interimAiText = "";
});
inputSelect.addEventListener("change", () => { inputSelect.addEventListener("change", () => {
if (micStream) { if (micStream) {
stopMic(); stopMic();

18
processors/vad.py
View File

@@ -63,6 +63,7 @@ class SileroVAD:
self.min_chunk_size = 512 self.min_chunk_size = 512
self.last_label = "Silence" self.last_label = "Silence"
self.last_probability = 0.0 self.last_probability = 0.0
self._energy_noise_floor = 1e-4
def _reset_state(self): def _reset_state(self):
# Silero VAD V4+ expects state shape [2, 1, 128] # Silero VAD V4+ expects state shape [2, 1, 128]
@@ -81,8 +82,7 @@ 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:
# Fallback energy-based VAD when model isn't available. # Fallback energy-based VAD with adaptive noise floor.
# Map RMS energy to a pseudo-probability so the existing threshold works.
if not pcm_bytes: if not pcm_bytes:
return "Silence", 0.0 return "Silence", 0.0
audio_int16 = np.frombuffer(pcm_bytes, dtype=np.int16) audio_int16 = np.frombuffer(pcm_bytes, dtype=np.int16)
@@ -90,9 +90,17 @@ class SileroVAD:
return "Silence", 0.0 return "Silence", 0.0
audio_float = audio_int16.astype(np.float32) / 32768.0 audio_float = audio_int16.astype(np.float32) / 32768.0
rms = float(np.sqrt(np.mean(audio_float * audio_float))) 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. # Update adaptive noise floor (slowly rises, faster to fall)
probability = min(1.0, rms / 0.05) if rms < self._energy_noise_floor:
self._energy_noise_floor = 0.95 * self._energy_noise_floor + 0.05 * rms
else:
self._energy_noise_floor = 0.995 * self._energy_noise_floor + 0.005 * rms
# Compute SNR-like ratio and map to probability
denom = max(self._energy_noise_floor, 1e-6)
snr = max(0.0, (rms - denom) / denom)
probability = min(1.0, snr / 3.0) # ~3x above noise => strong speech
label = "Speech" if probability >= 0.5 else "Silence" label = "Speech" if probability >= 0.5 else "Silence"
return label, probability return label, probability