Initial commit — voice pipeline experiment

STT (Silero VAD + Whisper via sherpa-onnx), Chatterbox TTS HTTP server,
query completeness classifier (Ollama), multi-voice demo scripts, and
planning docs. Kept as reference; clean rewrite planned in separate repos.

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

demo-kokoro.mjs

@@ -0,0 +1,85 @@
+/**
+ * Voice demo: microphone → Whisper STT → (optional LLM cleanup) → Kokoro TTS
+ *
+ * Environment variables:
+ *   WHISPER_MODEL   default: base.en
+ *                   options: tiny.en  base.en  small.en  medium.en  large-v3  large-v3-turbo
+ *   VOICE           default: af_heart
+ *                   options: af_heart af_bella af_nicole af_sarah af_sky am_adam am_michael
+ *   STT_PROVIDER    default: cuda
+ *   USE_LLM         set to 0 to disable (default: 1)
+ *   OLLAMA_MODEL    default: phi3:mini
+ */
+
+import { Stt }                             from './lib/stt.mjs'
+import { Tts, VOICES }                     from './lib/tts.mjs'
+import { llm_available, list_models, cleanup } from './lib/llm.mjs'
+
+const WHISPER_MODEL = process.env.WHISPER_MODEL || 'base.en'
+const VOICE         = process.env.VOICE         || 'af_heart'
+const STT_PROVIDER  = process.env.STT_PROVIDER  || 'cuda'
+const USE_LLM       = process.env.USE_LLM !== '0'
+
+if (!(VOICE in VOICES)) {
+	process.stderr.write(`[voice] unknown voice "${VOICE}". Available: ${Object.keys(VOICES).join(', ')}\n`)
+	process.exit(1)
+}
+
+process.stderr.write(`[stt] loading whisper-${WHISPER_MODEL} on ${STT_PROVIDER}...\n`)
+const stt = new Stt({ whisper_name: WHISPER_MODEL, provider: STT_PROVIDER })
+stt.init()
+process.stderr.write('[stt] ready\n')
+
+process.stderr.write('[tts] loading Kokoro...\n')
+const tts = new Tts({ voice: VOICE })
+tts.init()
+process.stderr.write('[tts] ready\n')
+
+let has_llm = false
+if (USE_LLM) {
+	has_llm = await llm_available()
+	if (has_llm) {
+		const models = await list_models()
+		process.stderr.write(`[llm] ollama available. models: ${models.slice(0, 6).join(', ')}\n`)
+	} else {
+		process.stderr.write('[llm] ollama not reachable, skipping cleanup\n')
+	}
+}
+
+process.stderr.write(`\n[ready] whisper=${WHISPER_MODEL} voice=${VOICE} llm=${has_llm}\n`)
+process.stderr.write('[ready] speak into your microphone. Ctrl+C to stop.\n\n')
+
+let speaking = false
+
+const stop = stt.listen(async (raw_text) => {
+	if (speaking) {
+		process.stderr.write(`[skipped] ${raw_text}\n`)
+		return
+	}
+	speaking = true
+
+	try {
+		process.stdout.write(`[raw]  ${raw_text}\n`)
+
+		let text = raw_text
+		if (has_llm) {
+			text = await cleanup(raw_text)
+			if (text !== raw_text) {
+				process.stdout.write(`[llm]  ${text}\n`)
+			}
+		}
+
+		await tts.speak_streaming(text)
+		process.stdout.write('\n')
+	} catch (err) {
+		process.stderr.write(`[error] ${err.message}\n`)
+	} finally {
+		speaking = false
+	}
+})
+
+process.on('SIGINT', () => {
+	process.stderr.write('\n[voice] stopping\n')
+	stop()
+	process.exit(0)
+})