WeLe Agentic AI with Docker deployment

src/speech/models.js

@@ -0,0 +1,118 @@
+// ============================================
+// Speech models — all ONNX, all CPU, all in this Node process.
+//
+// There is no GPU and no Python. That is the whole point: the AWS host has
+// neither, and a second service was one more thing to deploy and keep alive.
+//
+//   VAD  Silero            2 MB   endpointing
+//   STT  Whisper base     ~80 MB  Tamil + English + language detection
+//   TTS  MMS-TTS VITS    ~114 MB  per language, feed-forward
+//
+// Measured on an i7-10850H, CPU only:
+//   TTS  RTF 0.28x  (3.5x faster than realtime)
+//   STT  ~1.2 s for 4 s of audio
+//
+// Two findings worth keeping:
+//
+// * VITS is feed-forward. The earlier Parler-TTS attempt was autoregressive
+//   and ran at RTF ~5x — i.e. 5x SLOWER than realtime — which is why voice was
+//   unusable even on a GPU. Architecture mattered far more than hardware here.
+//
+// * int8 is a trap for a model this small: dynamic quantisation made TTS 5.7x
+//   SLOWER than fp32 (RTF 1.67x vs 0.28x) because the quantise/dequantise
+//   overhead dominates. We ship fp32 deliberately.
+// ============================================
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { pipeline, env } from '@huggingface/transformers';
+import logger from '../utils/logger.js';
+
+const ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../..');
+
+// Hub downloads are cached here so a container restart does not re-fetch.
+env.cacheDir = process.env.SPEECH_CACHE_DIR || path.join(ROOT, '.transformers-cache');
+// Tamil is loaded from a folder we exported ourselves — no public ONNX build
+// of mms-tts-tam exists. See scripts/export-tamil-tts.py.
+env.localModelPath = path.join(ROOT, 'assets/tts');
+
+const STT_MODEL = process.env.STT_MODEL || 'onnx-community/whisper-base';
+
+/** TTS voice per language. Tamil is local; English comes from the Hub. */
+const VOICES = {
+  ta: { id: 'mms-tts-tam', local: true },
+  en: { id: 'Xenova/mms-tts-eng', local: false },
+};
+
+export const LANGUAGES = [
+  { code: 'auto', label: 'Auto-detect', native: 'Auto' },
+  { code: 'ta', label: 'Tamil', native: 'தமிழ்' },
+  { code: 'en', label: 'English', native: 'English' },
+];
+
+const cache = new Map();
+let sttPromise = null;
+
+/**
+ * Models load on first use, not at boot. A CRM restart should not wait ~10 s
+ * for speech models that most sessions never touch.
+ */
+async function loadOnce(key, build) {
+  if (!cache.has(key)) {
+    const t0 = Date.now();
+    cache.set(key, build().then((m) => {
+      logger.info(`🔊 loaded ${key} in ${((Date.now() - t0) / 1000).toFixed(1)}s`);
+      return m;
+    }).catch((e) => {
+      cache.delete(key);              // let the next attempt retry
+      throw e;
+    }));
+  }
+  return cache.get(key);
+}
+
+export async function getSTT() {
+  if (!sttPromise) {
+    sttPromise = loadOnce(STT_MODEL, () =>
+      // q8 is the right call for Whisper — unlike VITS it is big enough that
+      // quantisation is a clear win.
+      pipeline('automatic-speech-recognition', STT_MODEL, { dtype: 'q8' }),
+    ).catch((e) => { sttPromise = null; throw e; });
+  }
+  return sttPromise;
+}
+
+export async function getTTS(lang = 'ta') {
+  const voice = VOICES[lang] || VOICES.ta;
+  const prev = env.allowRemoteModels;
+  try {
+    // Local folders must not be looked up on the Hub, and vice versa.
+    env.allowRemoteModels = !voice.local;
+    return await loadOnce(`tts:${voice.id}`, () =>
+      pipeline('text-to-speech', voice.id, { dtype: 'fp32' }),
+    );
+  } finally {
+    env.allowRemoteModels = prev;
+  }
+}
+
+export const supportsTTS = (lang) => Boolean(VOICES[lang]);
+
+/** Warm the models the deployment actually expects to use. */
+export async function warmup(langs = ['ta', 'en']) {
+  try {
+    await getSTT();
+    for (const l of langs) await getTTS(l);
+    logger.info('🔊 speech models warm');
+  } catch (e) {
+    logger.warn(`speech warmup failed (will retry on first use): ${e.message}`);
+  }
+}
+
+export function speechStatus() {
+  return {
+    stt_model: STT_MODEL,
+    tts_voices: Object.fromEntries(Object.entries(VOICES).map(([k, v]) => [k, v.id])),
+    loaded: [...cache.keys()],
+    languages: LANGUAGES,
+  };
+}