From d62fcdd1cd7aceb0a8a824a9fdbfd44423e21b11 Mon Sep 17 00:00:00 2001
From: Wong Ding Feng <dingfengwong@gmail.com>
Date: Fri, 29 May 2026 02:45:09 +0800
Subject: [PATCH] feat: add silence calibration and VAD state machine

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 transcribe.py | 71 +++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 71 insertions(+)

diff --git a/transcribe.py b/transcribe.py
index efa7f0c..e935be2 100644
--- a/transcribe.py
+++ b/transcribe.py
@@ -66,6 +66,77 @@ def main():
         print(f"\nTranscription:\n{text}\n")
 
 
+def calibrate_silence(duration=0.5):
+    print("Calibrating silence threshold...")
+    audio = sd.rec(int(duration * SAMPLE_RATE), samplerate=SAMPLE_RATE, channels=1, dtype="float32")
+    sd.wait()
+    rms = np.sqrt(np.mean(audio ** 2))
+    threshold = max(rms * 3, 0.01)
+    print(f"  Ambient RMS: {rms:.4f}, threshold: {threshold:.4f}")
+    return threshold
+
+
+FRAME_SIZE = 800          # 50ms at 16kHz
+PRE_ROLL_FRAMES = 6       # ~0.3s of audio before speech onset
+SILENCE_FRAMES = 16       # ~0.8s of silence to end a segment
+SPEECH_ONSET_FRAMES = 3   # ~150ms of speech to trigger
+MAX_SPEECH_SECONDS = 30   # force chunk boundary
+
+
+class VADStateMachine:
+    def __init__(self, threshold):
+        self.threshold = threshold
+        self.speaking = False
+        self.speech_frames = 0
+        self.silence_frames = 0
+        self.pre_roll = []
+        self.segment = []
+        self.segment_start_time = 0.0
+
+    def process_frame(self, frame, elapsed_time):
+        """Process one 50ms frame. Returns a (start_time, audio_array) tuple when a
+        complete speech segment is detected, otherwise None."""
+        rms = np.sqrt(np.mean(frame ** 2))
+        is_loud = rms > self.threshold
+
+        if not self.speaking:
+            self.pre_roll.append(frame)
+            if len(self.pre_roll) > PRE_ROLL_FRAMES:
+                self.pre_roll.pop(0)
+
+            if is_loud:
+                self.speech_frames += 1
+                if self.speech_frames >= SPEECH_ONSET_FRAMES:
+                    self.speaking = True
+                    self.silence_frames = 0
+                    self.segment = list(self.pre_roll)
+                    self.segment_start_time = max(0.0, elapsed_time - len(self.pre_roll) * FRAME_SIZE / SAMPLE_RATE)
+                    self.pre_roll = []
+            else:
+                self.speech_frames = 0
+            return None
+
+        # Currently speaking
+        self.segment.append(frame)
+
+        if is_loud:
+            self.silence_frames = 0
+        else:
+            self.silence_frames += 1
+
+        segment_duration = len(self.segment) * FRAME_SIZE / SAMPLE_RATE
+        if self.silence_frames >= SILENCE_FRAMES or segment_duration >= MAX_SPEECH_SECONDS:
+            result = (self.segment_start_time, np.concatenate(self.segment))
+            self.speaking = False
+            self.speech_frames = 0
+            self.silence_frames = 0
+            self.segment = []
+            self.pre_roll = []
+            return result
+
+        return None
+
+
 def stream_transcribe(processor, model, language):
     print("TODO: streaming mode")