Debounce speech detection — require 3 consecutive loud chunks (~150ms)

A single transient noise no longer resets the silence timer.
Only sustained audio energy counts as speech.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

query-demo.mjs

@@ -81,8 +81,10 @@ async function submit(query) {
 
 const SEND_WORDS = new Set(['go', 'done', 'send'])
 
-let accumulated    = ''
-let silence_timer  = null
+let accumulated         = ''
+let silence_timer       = null
+let loud_chunk_streak   = 0
+const LOUD_STREAK_NEEDED = 3   // ~150ms of sustained loud audio to count as speech
 
 function reset_silence_timer() {
 	clearTimeout(silence_timer)
@@ -123,10 +125,14 @@ stt.listen(async (text) => {
 	await submit(query)
 }, { on_audio: (chunk) => {
 	if (!accumulated) return
-	// Only reset silence timer if audio has significant energy (not silence)
 	const samples = new Int16Array(chunk.buffer, chunk.byteOffset, chunk.byteLength >> 1)
 	let sum = 0
 	for (let i = 0; i < samples.length; i++) sum += samples[i] * samples[i]
 	const rms = Math.sqrt(sum / samples.length) / 32768
-	if (rms > 0.02) reset_silence_timer()
+	if (rms > 0.02) {
+		loud_chunk_streak++
+		if (loud_chunk_streak >= LOUD_STREAK_NEEDED) reset_silence_timer()
+	} else {
+		loud_chunk_streak = 0
+	}
 } })