WeLe Agentic AI with Docker deployment

src/speech/index.js

@@ -7,10 +7,10 @@
 // one shift and should never have to touch a language menu.
 // ============================================
 import { Tensor } from '@huggingface/transformers';
-import { getSTT, getTTS, supportsTTS } from './models.js';
+import { getSTT, getTTS, supportsTTS, sttIsEnglishOnly, defaultLanguage } from './models.js';
 import logger from '../utils/logger.js';
 
-export { LANGUAGES, warmup, speechStatus } from './models.js';
+export { LANGUAGES, warmup, speechStatus, defaultLanguage } from './models.js';
 export { Endpointer, warmupVad } from './vad.js';
 
 const RATE = 16000;
@@ -67,7 +67,7 @@ async function detectLanguage(audio) {
  * @param {string} lang         'auto' | 'ta' | 'en'
  * @param {string} prefer       used when detection is unusable
  */
-export async function transcribe(audio, lang = 'auto', prefer = 'ta') {
+export async function transcribe(audio, lang = 'auto', prefer = defaultLanguage()) {
   if (!audio || audio.length < RATE / 5) {              // under 200 ms
     return { text: '', lang: prefer, note: 'too short' };
   }
@@ -80,7 +80,11 @@ export async function transcribe(audio, lang = 'auto', prefer = 'ta') {
   let detected = null;
   let confidence = null;
 
-  if (lang === 'auto') {
+  // An English-only checkpoint has no language tokens to read, and passing
+  // `language` to it is rejected — so "auto" simply means English there.
+  if (lang === 'auto' && sttIsEnglishOnly()) {
+    used = 'en';
+  } else if (lang === 'auto') {
     try {
       const d = await detectLanguage(audio);
       detected = d.lang;
@@ -95,7 +99,15 @@ export async function transcribe(audio, lang = 'auto', prefer = 'ta') {
     }
   }
 
-  const result = await stt(audio, { task: 'transcribe', language: used, return_timestamps: false });
+  // An English-only checkpoint rejects BOTH `task` and `language` — it has no
+  // other mode to select. Multilingual builds require the language, since they
+  // silently default to English otherwise.
+  const opts = { return_timestamps: false };
+  if (!sttIsEnglishOnly()) {
+    opts.task = 'transcribe';
+    opts.language = used;
+  }
+  const result = await stt(audio, opts);
   return finish(result, used, audio, t0, detected, confidence);
 }
 
@@ -115,11 +127,11 @@ function finish(result, used, audio, t0, detected, confidence) {
  * Synthesise one piece of text.
  * @returns {Promise<{audio: Float32Array, sampling_rate: number}>}
  */
-export async function synthesize(text, lang = 'ta') {
+export async function synthesize(text, lang = defaultLanguage()) {
   const clean = (text || '').trim();
   if (!clean) return null;
 
-  const use = supportsTTS(lang) ? lang : 'en';
+  const use = supportsTTS(lang) ? lang : defaultLanguage();
   const tts = await getTTS(use);
 
   const t0 = Date.now();