simplify speech and embeddings through litellm

2026-05-09 05:09:02 +02:00 · 2026-05-09 05:09:02 +02:00 · 1b3ea569b7
commit 1b3ea569b7
parent 79037fa775
24 changed files with 184 additions and 1242 deletions
--- a/docker-compose.yml
+++ b/docker-compose.yml
@ -15,6 +15,7 @@ services:
      TTS_PROVIDER: litellm
      LITELLM_TTS_MODEL: local-kokoro-tts
      LITELLM_TTS_VOICE: sherpa/kokoro:am_adam
      LITELLM_TTS_VOICES: sherpa/kokoro:am_adam,sherpa/kokoro:am_michael,sherpa/kokoro:af_bella,sherpa/kokoro:af_nicole,sherpa/kokoro:bf_emma,sherpa/kokoro:bm_lewis
    volumes:
      - scribe-logs:/app/data/logs
      - clinical-assistant-mcp-data:/app/mcp-data:ro
--- a/docs/api-reference.md
+++ b/docs/api-reference.md
@ -591,7 +591,7 @@ Check whether speech-to-text transcription is available and which provider is co
  ```json
  {
    "available": true,
-    "provider": "google | aws | litellm | openai | local | none"
+    "provider": "litellm | none"
  }
  ```
--- a/docs/architecture.md
+++ b/docs/architecture.md
@ -48,8 +48,8 @@ src/
    logger.js                    # audit/api/access + Loki shipper
    errors.js                    # generic 500 responder
    models.js, prompts.js, ai.js # AI provider + model + prompt management
-    embeddings.js                # Vertex / LiteLLM / OpenAI embeddings
+    embeddings.js                # LiteLLM embeddings
-    transcribe*.js, tts*.js      # STT / TTS provider clients
+    transcribe.js, tts.js        # LiteLLM STT / TTS routes
  routes/                        # Express routers (auth, hpi, soap, patient education, …)
 public/                          # SPA
@ -163,4 +163,4 @@ The clinical assistant can call an external MCP-backed retrieval service. Ped-AI
 ## Speech
-Browser Whisper and browser-local Whisper model downloads are removed from runtime. Speech-to-text routes through configured server-side providers. Browser-native Web Speech remains available only when explicitly enabled by user settings and browser support.
+Browser Whisper and browser-local Whisper model downloads are removed from runtime. Speech-to-text routes through LiteLLM; upstream provider choice belongs in LiteLLM config. Browser-native Web Speech remains available only when explicitly enabled by user settings and browser support.
--- a/docs/configuration.md
+++ b/docs/configuration.md
@ -31,38 +31,31 @@ keys):
 |---|---|
 | `AI_PROVIDER` | `openrouter` / `bedrock` / `azure` / `vertex` / `litellm`. If unset, the startup loader uses configured credentials and the last initialized provider in Bedrock → Azure → Vertex → LiteLLM order wins; otherwise OpenRouter is the default. |
 | `OPENROUTER_API_KEY` | OpenRouter key (not HIPAA-eligible). |
-| `AWS_BEDROCK_REGION`, `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY` | Bedrock / Transcribe / Transcribe-Medical. |
+| `AWS_BEDROCK_REGION`, `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY` | Bedrock chat provider. |
 | `AZURE_OPENAI_ENDPOINT`, `AZURE_OPENAI_API_KEY`, `AZURE_DEPLOYMENT_NAME`, `AZURE_OPENAI_API_VERSION` | Azure OpenAI. |
-| `GOOGLE_VERTEX_PROJECT`, `GOOGLE_VERTEX_LOCATION`, `GOOGLE_APPLICATION_CREDENTIALS` | Vertex AI + Gemini chat/STT/TTS. |
+| `GOOGLE_VERTEX_PROJECT`, `GOOGLE_VERTEX_LOCATION`, `GOOGLE_APPLICATION_CREDENTIALS` | Vertex AI chat provider. |
 | `LITELLM_API_BASE`, `LITELLM_API_KEY` | OpenAI-compatible AI gateway (Bifrost, LiteLLM, or similar). |
 ### Speech-to-text
 | Variable | Purpose |
 |---|---|
-| `TRANSCRIBE_PROVIDER` | `google`, `aws`, `local`, `openai`, `litellm`. Auto-detects if unset. |
+| `TRANSCRIBE_PROVIDER` | Use `litellm`; auto mode uses LiteLLM when configured. |
 | `OPENAI_API_KEY` | OpenAI Whisper. |
 | `GOOGLE_STT_MODEL` | Gemini model used as STT (default `gemini-2.0-flash`). |
 | `AWS_TRANSCRIBE_MEDICAL` | `true` enables Transcribe Medical. |
 | `AWS_TRANSCRIBE_SPECIALTY` | `PRIMARYCARE` / `CARDIOLOGY` / `NEUROLOGY` / `ONCOLOGY` / `RADIOLOGY` / `UROLOGY`. |
 | `WHISPER_BINARY`, `WHISPER_MODEL_SIZE`, `WHISPER_MODEL_PATH`, `WHISPER_LANGUAGE`, `WHISPER_THREADS` | Local whisper.cpp / faster-whisper. |
 | `LITELLM_STT_MODEL` | Model name for LiteLLM-routed STT. |
 ### Text-to-speech
 | Variable | Purpose |
 |---|---|
-| `GOOGLE_TTS_VOICE` | Google Cloud TTS voice (e.g. `en-US-Journey-F`). |
+| `LITELLM_TTS_MODEL`, `LITELLM_TTS_VOICE` | LiteLLM-routed TTS model and default voice. |
-| `ELEVENLABS_API_KEY` | ElevenLabs (not HIPAA-compliant). |
+| `LITELLM_TTS_VOICES` | Comma-separated LiteLLM-compatible voices exposed in voice search and user preferences. |
 | `LITELLM_TTS_MODEL`, `LITELLM_TTS_VOICE` | LiteLLM-routed TTS. |
 ### Embeddings
 | Variable | Purpose |
 |---|---|
-| `EMBEDDING_MODEL` | Embedding model name (default `vertex_ai/text-embedding-005`). |
+| `EMBEDDING_MODEL` | LiteLLM embedding model name (default `openai-text-embedding-3-large`). |
-| `EMBEDDING_DIMENSIONS` | Vector dimensions (default 768). |
+| `EMBEDDING_DIMENSIONS` | Vector dimensions (default 3072). |
 | `VERTEX_PROJECT`, `GOOGLE_CLOUD_PROJECT` | Direct Vertex embedding project. Embedding utility also works through LiteLLM when `LITELLM_API_BASE` is set. |
 ### Email (SMTP)
--- a/docs/developer-guide.md
+++ b/docs/developer-guide.md
@ -37,9 +37,9 @@ src/
    fileType.js                  magic-byte upload verifier
    errors.js                    generic 500 responder
    logger.js                    audit + api + access + Loki shipper
-    embeddings.js                Vertex / LiteLLM / OpenAI embeddings
+    embeddings.js                LiteLLM embeddings
    notify.js                    ntfy push
-    transcribe*.js, tts*.js      STT / TTS provider clients
+    transcribe.js, tts.js        LiteLLM STT / TTS routes
  routes/                        Express routers for auth, AI workflows, education, logs, and user data
 public/
@ -272,8 +272,8 @@ docker exec -w /app pediatric-ai-scribe npm run migrate:new -- add_my_table
 | `sickVisit.js` | `/api` | Auth | Sick visit |
 | `milestones.js` | `/api` | Auth | Developmental milestone narratives |
 | `refine.js` | `/api` | Auth | Refine / shorten / clarify |
-| `transcribe.js` | `/api` | Auth | STT (5 providers) |
+| `transcribe.js` | `/api` | Auth | LiteLLM STT |
-| `tts.js` | `/api` | Auth | TTS (3 providers) |
+| `tts.js` | `/api` | Auth | LiteLLM TTS |
 | `encounters.js` | `/api` | Auth | Save / load / optimistic-lock encounters |
 | `memories.js` | `/api` | Auth | Templates + prompt preferences |
 | `audioBackups.js` | `/api` | Auth | Encrypted audio retry store |
--- a/docs/embeddings-setup.md
+++ b/docs/embeddings-setup.md
@ -10,7 +10,7 @@ This guide explains how to set up and use the new vector-based semantic search f
  - **Semantic** (`/api/learning/search/semantic`) - AI-powered vector similarity
  - **Hybrid** (`/api/learning/search/hybrid`) - Combines both for best results
 - **Auto-embedding** - Content is automatically vectorized when created/updated
- **HIPAA-compliant** - Uses Vertex AI embeddings (BAA available)
+- **Gateway-routed** - Uses LiteLLM embeddings so provider policy stays in one place
 ## Prerequisites
@ -37,37 +37,22 @@ postgres:
  # ... rest of your config
 ```
-### 2. Configure Embedding Provider
+### 2. Configure LiteLLM Embeddings
 Add to your `.env` file:
 ```bash
 # Option 1: Vertex AI (HIPAA-eligible, recommended)
 EMBEDDING_MODEL=vertex_ai/text-embedding-005
 EMBEDDING_DIMENSIONS=768
 VERTEX_PROJECT=your-gcp-project-id
 VERTEX_LOCATION=us-central1
 GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json
 # Option 2: LiteLLM Proxy (routes to any provider)
 LITELLM_API_BASE=http://localhost:4000
 LITELLM_API_KEY=your-key
-EMBEDDING_MODEL=text-embedding-005  # LiteLLM will route to configured provider
+EMBEDDING_MODEL=openai-text-embedding-3-large
-
+EMBEDDING_DIMENSIONS=3072
 # Option 3: OpenAI (NOT HIPAA-eligible, fallback only)
 OPENAI_API_KEY=sk-your-key
 # Uses text-embedding-3-small and truncates to the 768-dimension column
 ```
-## Available Vertex AI Embedding Models
+## Available Embedding Models
-Tested and working via LiteLLM:
+The Admin embedding search reads LiteLLM `/model/info` and only shows models with `model_info.mode = "embedding"`. Do not add app-side built-in Vertex/OpenAI embedding lists; configure those choices in LiteLLM.
-| Model | Dimensions | Use Case | HIPAA |
+The local LiteLLM instance currently exposes examples such as `openai-text-embedding-3-large`, `openai-text-embedding-3-small`, and Mistral embedding models. Dimensions are read from LiteLLM metadata when available.
 |-------|-----------|----------|-------|
 | **vertex_ai/text-embedding-005** | 768 | English + code (recommended) | Yes |
 | **vertex_ai/gemini-embedding-001** | 768-3072 | Multilingual + code, best quality | Yes |
 | **vertex_ai/text-multilingual-embedding-002** | 768 | Multilingual focus | Yes |
 ## Setup Steps
@ -113,8 +98,8 @@ Response:
  "total": 50,
  "withEmbeddings": 50,
  "missing": 0,
-  "model": "vertex_ai/text-embedding-005",
+  "model": "openai-text-embedding-3-large",
-  "dimensions": 768
+  "dimensions": 3072
 }
 ```
@ -148,8 +133,8 @@ Combines keyword + semantic for best results. Automatically deduplicates and ran
 1. **Content Creation/Update**:
   - Text is extracted from `title`, `subject`, and `body` (HTML stripped)
-   - Sent to embedding model (Vertex AI)
+   - Sent to the configured LiteLLM embedding model
-   - Returns 768-dimensional vector
+   - Returns an embedding vector
   - Stored in `learning_content.embedding` column
 2. **Semantic Search**:
@ -166,11 +151,7 @@ Combines keyword + semantic for best results. Automatically deduplicates and ran
 ## Cost Estimate
-**Google Vertex AI pricing:**
+Embedding cost depends on the upstream configured in LiteLLM.
 - text-embedding-005: $0.025 per 1M characters
 - Average article: 10,000 chars = $0.00025
 - 1,000 articles: ~**$0.25 one-time**
 - Search queries: ~$0.0000125 per query
 ## Troubleshooting
@ -179,14 +160,12 @@ Combines keyword + semantic for best results. Automatically deduplicates and ran
 - For Docker: Use `pgvector/pgvector:pg16` image
 ### "Embeddings not configured"
- Verify `.env` has `VERTEX_PROJECT` / `GOOGLE_CLOUD_PROJECT`, `LITELLM_API_BASE`, or `OPENAI_API_KEY`
+- Verify `.env` has `LITELLM_API_BASE`
 - Check service account credentials: `GOOGLE_APPLICATION_CREDENTIALS`
 - Test: `curl http://localhost:3000/api/admin/learning/embeddings/status`
 ### "Embedding generation failed"
 - Check logs for API errors
- Verify Vertex AI API is enabled in GCP
+- Verify LiteLLM `/model/info` shows the selected model with `mode: embedding`
 - Verify service account has `aiplatform.endpoints.predict` permission
 - Check content isn't empty (skips empty bodies)
 ### "No results from semantic search"
@ -196,7 +175,7 @@ Combines keyword + semantic for best results. Automatically deduplicates and ran
 ## Performance
- **Embedding generation**: ~500ms per article (Vertex AI)
+- **Embedding generation**: latency depends on the LiteLLM upstream
 - **Search latency**:
  - Keyword: 10-50ms
  - Semantic: 20-100ms (with IVFFLAT index)
@ -205,7 +184,7 @@ Combines keyword + semantic for best results. Automatically deduplicates and ran
 ## Security And Compliance
- **HIPAA-eligible**: Vertex AI supports BAA (Business Associate Agreement)
+- **Compliance**: controlled by the upstream provider configured in LiteLLM
 - **Data retention**: Embeddings stored in your database only
 - **No PHI**: Only article content (not patient data) is embedded
 - **Encryption**: TLS in transit, at-rest encryption via PostgreSQL
--- a/docs/features-explained.md
+++ b/docs/features-explained.md
@ -16,7 +16,7 @@ The bedside tools include a per-user phone extension and pager directory. Entrie
 ## Speech
-Final transcription is server-side. Configure Google/Gemini, AWS Transcribe, LiteLLM, or OpenAI Whisper according to your deployment requirements.
+Final transcription is server-side through LiteLLM. Configure upstream STT providers in LiteLLM rather than in Ped-AI.
 Browser-native Web Speech is only an explicit opt-in preview path. It is not the final clinical transcript and may use browser-vendor cloud services.
@ -24,7 +24,7 @@ Browser Whisper and browser-local model workers are removed. Do not expect a pre
 ## Text To Speech
-The voice preview button calls the configured TTS provider and plays the returned audio in the browser. If preview is silent, check that a voice is selected, a provider is configured, the user is authenticated, and browser autoplay has not blocked playback.
+The voice preview button calls LiteLLM TTS and plays the returned audio in the browser. If preview is silent, check that a LiteLLM voice is selected, the gateway is configured, the user is authenticated, and browser autoplay has not blocked playback.
 ## Learning Hub
--- a/docs/logic/ai-and-voice.md
+++ b/docs/logic/ai-and-voice.md
@ -54,11 +54,10 @@ as low-priority style/template context. `custom` memories and legacy
 `POST /api/transcribe` accepts one audio file up to 25 MB. Provider selection:
- explicit `TRANSCRIBE_PROVIDER=google|aws|local|openai|litellm`, or
+- explicit `TRANSCRIBE_PROVIDER=litellm`, or
- auto mode: Google/Gemini, then AWS Transcribe, then direct OpenAI Whisper.
+- auto mode when `LITELLM_API_BASE` is configured.
-LiteLLM and local STT are explicit choices. Browser Whisper/browser-local model
+Direct Google, AWS, local Whisper, and OpenAI Whisper branches are not part of the runtime. Browser Whisper/browser-local model downloads are also absent.
 downloads are not part of the runtime.
 ## Browser Web Speech
@ -75,9 +74,9 @@ store the failed recording.
 ## Text-To-Speech
-`POST /api/text-to-speech` returns `audio/mpeg` and marks the provider in
+`POST /api/text-to-speech` returns audio from LiteLLM and marks the LiteLLM
-`X-TTS-Provider`. Supported providers include Google Cloud TTS, LiteLLM-routed
+model in `X-TTS-Provider`. Voices are LiteLLM-compatible strings configured by
-TTS, and ElevenLabs when configured.
+`LITELLM_TTS_VOICES`.
 ## Post-Note Helpers
--- a/docs/speech.md
+++ b/docs/speech.md
@ -2,16 +2,13 @@
 ## Transcription
-`POST /api/transcribe` accepts `multipart/form-data` with one audio file up to 25 MB. The provider is selected by `TRANSCRIBE_PROVIDER`, or auto-detected from available credentials.
+`POST /api/transcribe` accepts `multipart/form-data` with one audio file up to 25 MB. Server STT is routed through LiteLLM.
-Provider priority in auto mode is Google/Gemini, then AWS Transcribe, then direct OpenAI Whisper. Set `TRANSCRIBE_PROVIDER=litellm` or `TRANSCRIBE_PROVIDER=local` explicitly for those paths.
+Set `TRANSCRIBE_PROVIDER=litellm`, `LITELLM_API_BASE`, and `LITELLM_STT_MODEL`. Auto mode also uses LiteLLM when the gateway is configured.
 | Provider | Notes | HIPAA posture |
 |---|---|---|
 | Google/Gemini | Uses the configured Vertex/Gemini STT model. | Eligible with the correct Google Cloud agreement. |
 | AWS Transcribe | Supports standard and Medical mode. | Eligible with the correct AWS agreement. |
 | LiteLLM | Sends audio through the configured LiteLLM `/audio/transcriptions` backend. | Depends on the selected upstream. |
 | OpenAI Whisper | Uses `whisper-1` directly when `OPENAI_API_KEY` is configured. | Not HIPAA eligible unless your own agreement says otherwise. |
 Browser Whisper and browser-local Whisper workers are not part of the runtime. Do not add browser model downloads or Transformers.js STT back into the public app.
@ -21,13 +18,13 @@ Browser-native Web Speech can show interim text when the user explicitly enables
 ## Text To Speech
-`POST /api/text-to-speech` returns `audio/mpeg`. The `X-TTS-Provider` response header identifies the provider used. Requests are limited to 5000 characters.
+`POST /api/text-to-speech` returns audio from LiteLLM `/audio/speech`. The `X-TTS-Provider` response header identifies the LiteLLM model used. Requests are limited to 5000 characters.
 | Provider | Notes |
 |---|---|
 | Google Cloud TTS | Uses Google Cloud voices when configured. |
 | LiteLLM | Uses `LITELLM_TTS_MODEL` and `LITELLM_TTS_VOICE`. |
-| ElevenLabs | Available when configured; not HIPAA eligible by default. |
+
 The admin/user voice pickers read available LiteLLM-compatible voices from `LITELLM_TTS_VOICES`.
 ## Audio Backup
--- a/docs/transcription-options.md
+++ b/docs/transcription-options.md
@ -1,21 +1,17 @@
 # Transcription Options
-Ped-AI currently supports server-side transcription plus an explicit browser Web Speech preview option. Browser Whisper was removed and should not be offered in settings, documentation, public workers, or model download scripts.
+Ped-AI currently supports server-side transcription through LiteLLM plus an explicit browser Web Speech preview option. Browser Whisper was removed and should not be offered in settings, documentation, public workers, or model download scripts.
 ## Recommended Clinical Setup
-Use a server-side provider covered by your compliance requirements.
+Route STT through LiteLLM and configure the compliant upstream in LiteLLM.
 | Need | Recommended provider |
 |---|---|
-| HIPAA-eligible cloud STT | Google/Gemini through Vertex AI or AWS Transcribe with a BAA. |
+| Server STT | LiteLLM with a compliant upstream. |
 | OpenAI-compatible routing | LiteLLM with a compliant upstream. |
 | Direct OpenAI Whisper | Only when acceptable for your deployment. |
 | Real-time draft preview | Browser Web Speech only with explicit user opt-in and privacy warning. |
-Auto-detect mode currently checks Google/Gemini first, then AWS Transcribe,
+Auto-detect uses LiteLLM when `LITELLM_API_BASE` is configured. Direct Google, AWS, local Whisper, and OpenAI Whisper branches are not part of the app runtime.
 then direct OpenAI Whisper. Use `TRANSCRIBE_PROVIDER=litellm` or
 `TRANSCRIBE_PROVIDER=local` when you want those providers.
 ## Configuration
@ -23,26 +19,7 @@ then direct OpenAI Whisper. Use `TRANSCRIBE_PROVIDER=litellm` or
 TRANSCRIBE_PROVIDER=litellm
 LITELLM_API_BASE=https://your-litellm.example/v1
 LITELLM_API_KEY=<key>
-LITELLM_STT_MODEL=whisper-1
+LITELLM_STT_MODEL=local-parakeet-v3
 ```
 Other provider examples:
 ```env
 # Google/Gemini
 TRANSCRIBE_PROVIDER=google
 GOOGLE_VERTEX_PROJECT=your-project-id
 GOOGLE_STT_MODEL=gemini-2.0-flash
 # AWS Transcribe
 TRANSCRIBE_PROVIDER=aws
 AWS_BEDROCK_REGION=us-east-1
 AWS_TRANSCRIBE_MEDICAL=true
 AWS_TRANSCRIBE_SPECIALTY=PRIMARYCARE
 # Direct OpenAI Whisper
 TRANSCRIBE_PROVIDER=openai
 OPENAI_API_KEY=<key>
 ```
 ## Failure Handling
--- a/src/routes/adminConfig.js
+++ b/src/routes/adminConfig.js
@ -9,7 +9,7 @@ var { authMiddleware, adminMiddleware } = require('../middleware/auth');
 var PROMPTS = require('../utils/prompts');
 var logger = require('../utils/logger');
 var { gatewayUrl } = require('../utils/errors');
-var { GOOGLE_TTS_VOICES, getTTSEnvProvider, getLiteLLMTTSDiscoveryItems, getLiteLLMTTSModels, getTTSProvider, getTTSVoiceLists } = require('../utils/ttsProvider');
+var { getTTSEnvProvider, getLiteLLMTTSDiscoveryItems, getTTSProvider, getTTSVoiceLists } = require('../utils/ttsProvider');
 var { getLiteLLMHeaders } = require('../utils/litellm');
 var { getSTTDependencies, getLiteLLMSTTModels, getSTTModelLists, getSTTProvider } = require('../utils/sttProvider');
 var { getLiteLLMEmbeddingModels } = require('../utils/embeddings');
@ -461,12 +461,11 @@ router.post('/config/image-models/test', async function(req, res) {
 // ── GET TTS provider status, voice list, and DB overrides ────────────────
 router.get('/config/tts', async function(req, res) {
  try {
    var { isGoogleTTSConfigured } = require('../utils/ttsGoogle');
    var envProvider = getTTSEnvProvider();
    var activeProvider = getTTSProvider();
    var dbVoice = await db.getSetting('tts.voice') || '';
    var dbModel = await db.getSetting('tts.model') || '';
-    var envVoice = process.env.GOOGLE_TTS_VOICE || process.env.LITELLM_TTS_VOICE || '';
+    var envVoice = process.env.LITELLM_TTS_VOICE || '';
    var envModel = process.env.LITELLM_TTS_MODEL || '';
    res.json({
      success: true,
@ -479,9 +478,7 @@ router.get('/config/tts', async function(req, res) {
      envVoice: envVoice,
      envModel: envModel,
      configured: {
-        google: false,
+        litellm: !!process.env.LITELLM_API_BASE
        litellm: !!process.env.LITELLM_API_BASE,
        elevenlabs: false
      },
      voices: getTTSVoiceLists()
    });
@ -497,28 +494,6 @@ router.get('/config/tts/discover', async function(req, res) {
    var provider = getTTSProvider();
    if (provider === 'google') {
      try {
        var GoogleAuth = require('google-auth-library').GoogleAuth;
        var auth = new GoogleAuth({ scopes: ['https://www.googleapis.com/auth/cloud-platform'] });
        var client = await auth.getClient();
        var tokenData = await client.getAccessToken();
        var gResp = await axios.get('https://texttospeech.googleapis.com/v1/voices?languageCode=en', {
          headers: { 'Authorization': 'Bearer ' + tokenData.token }, timeout: 10000
        });
        if (gResp.data && gResp.data.voices) {
          gResp.data.voices.forEach(function(v) {
            discovered.push({ id: v.name, name: v.name + ' (' + (v.ssmlGender || 'NEUTRAL') + ')', gender: v.ssmlGender, source: 'google-api' });
          });
        }
      } catch (e) {
        logger.warn('Google TTS voice API failed: ' + e.message + ' — using built-in list');
        GOOGLE_TTS_VOICES.forEach(function(v) {
          discovered.push({ id: v, name: v, source: 'google-builtin' });
        });
      }
    }
    if (provider === 'litellm' && process.env.LITELLM_API_BASE) {
      var dbVoice = await db.getSetting('tts.voice') || '';
      var dbModel = await db.getSetting('tts.model') || '';
@ -535,19 +510,6 @@ router.get('/config/tts/discover', async function(req, res) {
      });
    }
    if (provider === 'elevenlabs' && process.env.ELEVENLABS_API_KEY) {
      try {
        var elResp = await axios.get('https://api.elevenlabs.io/v1/voices', {
          headers: { 'xi-api-key': process.env.ELEVENLABS_API_KEY }, timeout: 10000
        });
        if (elResp.data && elResp.data.voices) {
          elResp.data.voices.forEach(function(v) {
            discovered.push({ id: v.voice_id, name: v.name, source: 'elevenlabs-api' });
          });
        }
      } catch (e) { logger.warn('ElevenLabs voice list failed: ' + e.message); }
    }
    if (search) {
      discovered = discovered.filter(function(d) {
        return d.id.toLowerCase().indexOf(search) !== -1 || d.name.toLowerCase().indexOf(search) !== -1;
@ -562,53 +524,24 @@ router.post('/config/tts/test', async function(req, res) {
  try {
    var text = ((req.body.text || 'Hello, this is a TTS test for Pediatric AI Scribe.')).substring(0, 500);
    var voice = req.body.voice;
    var { synthesizeWithGoogleTTS } = require('../utils/ttsGoogle');
    var axios = require('axios');
    var provider = getTTSProvider();
    if (provider === 'none') return res.json({ success: false, error: 'No TTS provider configured' });
    if (provider !== 'litellm') return res.json({ success: false, error: 'TTS is configured for LiteLLM only' });
    if (!process.env.LITELLM_API_BASE) return res.json({ success: false, error: 'LITELLM_API_BASE not set' });
-    var buffer, usedVoice = voice;
+    var adminModel = await db.getSetting('tts.model') || '';
-    if (provider === 'google') {
+    var adminVoice = await db.getSetting('tts.voice') || '';
-      usedVoice = voice || process.env.GOOGLE_TTS_VOICE || 'en-US-Journey-F';
+    var ttsModel = adminModel || process.env.LITELLM_TTS_MODEL || '';
-      buffer = await synthesizeWithGoogleTTS(text, usedVoice);
+    var usedVoice = voice || adminVoice || process.env.LITELLM_TTS_VOICE || '';
-    } else if (provider === 'litellm') {
+    if (!ttsModel) return res.json({ success: false, error: 'No LiteLLM TTS model configured' });
-      if (!process.env.LITELLM_API_BASE) return res.json({ success: false, error: 'LITELLM_API_BASE not set' });
+
-      var db = require('../db/database');
+    var ttsResp = await axios.post(gatewayUrl('/audio/speech'),
-      var adminModel = await db.getSetting('tts.model') || '';
+      { model: ttsModel, voice: usedVoice, input: text },
-      var adminVoice = await db.getSetting('tts.voice') || '';
+      { headers: getLiteLLMHeaders('application/json'), responseType: 'arraybuffer', timeout: 60000 }
-      if (!adminModel && adminVoice) {
+    );
-        try {
+    var buffer = Buffer.from(ttsResp.data);
          var liteLLMBase = (process.env.LITELLM_API_BASE || '').replace(/\/+$/, '').replace(/\/v1\/?$/, '');
          var modelInfoResp = await axios.get(liteLLMBase + '/model/info', { headers: getLiteLLMHeaders(), timeout: 10000 });
          if (getLiteLLMTTSModels(modelInfoResp.data && modelInfoResp.data.data).indexOf(adminVoice) !== -1) {
            adminModel = adminVoice;
            adminVoice = '';
          }
        } catch (e) { logger.warn('LiteLLM TTS mode lookup failed: ' + e.message); }
      }
      var resolvedVoice = voice || adminVoice || '';
      var rawModel = adminModel || process.env.LITELLM_TTS_MODEL || 'tts-1';
      var ttsModel = rawModel;
      usedVoice = resolvedVoice || process.env.LITELLM_TTS_VOICE || 'alloy';
      // base URL handled by gatewayUrl() helper
      var ttsResp = await axios.post(gatewayUrl('/audio/speech'),
        { model: ttsModel, voice: usedVoice, input: text },
        { headers: getLiteLLMHeaders('application/json'), responseType: 'arraybuffer', timeout: 60000 }
      );
      buffer = Buffer.from(ttsResp.data);
    } else if (provider === 'elevenlabs') {
      if (!process.env.ELEVENLABS_API_KEY) return res.json({ success: false, error: 'ElevenLabs not configured' });
      usedVoice = voice || 'pNInz6obpgDQGcFmaJgB';
      var elResp = await axios({
        method: 'POST',
        url: 'https://api.elevenlabs.io/v1/text-to-speech/' + usedVoice,
        headers: { 'xi-api-key': process.env.ELEVENLABS_API_KEY, 'Content-Type': 'application/json' },
        data: { text: text, model_id: 'eleven_turbo_v2_5', voice_settings: { stability: 0.5, similarity_boost: 0.75 } },
        responseType: 'arraybuffer'
      });
      buffer = Buffer.from(elResp.data);
    }
    res.json({ success: true, audio: buffer.toString('base64'), provider: provider, voice: usedVoice });
  } catch (e) {
    var detail = e.response && e.response.data
@ -625,7 +558,7 @@ router.get('/config/stt', async function(req, res) {
    var configured = getSTTDependencies();
    var activeProvider = getSTTProvider(configured);
    var dbModel   = await db.getSetting('stt.model') || '';
-    var envModel  = process.env.GOOGLE_STT_MODEL || process.env.LITELLM_STT_MODEL || '';
+    var envModel  = process.env.LITELLM_STT_MODEL || '';
    res.json({
      success: true,
      provider: activeProvider,
@ -644,36 +577,10 @@ router.get('/config/stt/discover', async function(req, res) {
  try {
    var search = (req.query.q || '').toLowerCase().trim();
    var axios = require('axios');
    var { whisperClient } = require('../utils/ai');
    var discovered = [];
    var provider = getSTTProvider();
    if (provider === 'google') {
      ['gemini-2.5-flash','gemini-2.0-flash','gemini-2.0-flash-lite','gemini-1.5-pro','gemini-1.5-flash'].forEach(function(m) {
        discovered.push({ id: m, name: m, source: 'google-builtin' });
      });
    }
    if (provider === 'aws') {
      ['standard','medical'].forEach(function(m) {
        discovered.push({ id: m, name: 'AWS Transcribe ' + m.charAt(0).toUpperCase() + m.slice(1), source: 'aws-builtin' });
      });
    }
    if (provider === 'openai' && whisperClient) {
      try {
        var oModels = await whisperClient.models.list();
        oModels.data.forEach(function(m) {
          if (m.id.toLowerCase().indexOf('whisper') !== -1 || m.id.toLowerCase().indexOf('transcri') !== -1) {
            discovered.push({ id: m.id, name: m.id, source: 'openai-api' });
          }
        });
      } catch(e) {
        discovered.push({ id: 'whisper-1', name: 'whisper-1', source: 'openai-builtin' });
      }
    }
    if (provider === 'litellm' && process.env.LITELLM_API_BASE) {
      try {
        var liteLLMBase = (process.env.LITELLM_API_BASE || '').replace(/\/+$/, '').replace(/\/v1\/?$/, '');
@ -689,14 +596,6 @@ router.get('/config/stt/discover', async function(req, res) {
      }
    }
    if (provider === 'local') {
      var localModel = process.env.WHISPER_MODEL_SIZE || 'base';
      discovered.push({ id: localModel, name: 'Local Whisper: ' + localModel + ' (configured)', source: 'local-env' });
      ['tiny','base','small','medium','large'].forEach(function(m) {
        if (m !== localModel) discovered.push({ id: m, name: 'Local Whisper: ' + m, source: 'local-builtin' });
      });
    }
    if (search) {
      discovered = discovered.filter(function(d) {
        return d.id.toLowerCase().indexOf(search) !== -1 || d.name.toLowerCase().indexOf(search) !== -1;
@ -714,68 +613,27 @@ router.post('/config/stt/test', async function(req, res) {
    var audioBuffer = Buffer.from(audioBase64, 'base64');
    var mime = mimeType || 'audio/webm';
    var { transcribeWithGemini } = require('../utils/transcribeGoogle');
    var { transcribeWithAWS } = require('../utils/transcribeAWS');
    var { transcribeWithLocal } = require('../utils/transcribeLocal');
    var { whisperClient } = require('../utils/ai');
    var axios = require('axios');
    var provider = getSTTProvider();
    if (provider === 'none') return res.json({ success: false, error: 'No STT provider configured' });
    if (provider !== 'litellm') return res.json({ success: false, error: 'STT is configured for LiteLLM only' });
    if (!process.env.LITELLM_API_BASE) return res.json({ success: false, error: 'LITELLM_API_BASE not set' });
    var start = Date.now();
    var text = '';
-    if (provider === 'google') {
+    var adminSttModel = await db.getSetting('stt.model') || '';
-      var model = process.env.GOOGLE_STT_MODEL || 'gemini-2.0-flash';
+    var sttModel = adminSttModel || process.env.LITELLM_STT_MODEL || '';
-      text = await transcribeWithGemini(audioBuffer, mime, model);
+    if (!sttModel) return res.json({ success: false, error: 'No LiteLLM STT model configured' });
    } else if (provider === 'aws') {
      text = await transcribeWithAWS(audioBuffer, mime);
    } else if (provider === 'local') {
      text = await transcribeWithLocal(audioBuffer, mime);
    } else if (provider === 'litellm') {
      if (!process.env.LITELLM_API_BASE) return res.json({ success: false, error: 'LITELLM_API_BASE not set' });
      var db = require('../db/database');
      var adminSttModel = await db.getSetting('stt.model') || '';
      var sttModel = adminSttModel || process.env.LITELLM_STT_MODEL || 'gemini-2.0-flash';
-      var isTranscriptionModel = false;
+    var ext = mime.split('/')[1] || 'webm';
-      try {
+    var file = new File([audioBuffer], 'audio.' + ext, { type: mime });
-        var liteLLMBase = (process.env.LITELLM_API_BASE || '').replace(/\/+$/, '').replace(/\/v1\/?$/, '');
+    var form = new FormData();
-        var modelInfoResp = await axios.get(liteLLMBase + '/model/info', { headers: getLiteLLMHeaders(), timeout: 10000 });
+    form.append('file', file);
-        isTranscriptionModel = getLiteLLMSTTModels(modelInfoResp.data && modelInfoResp.data.data).indexOf(sttModel) !== -1;
+    form.append('model', sttModel);
-      } catch (e) { logger.warn('LiteLLM STT mode lookup failed: ' + e.message); }
+    var sttResp = await fetch(gatewayUrl('/audio/transcriptions'), {
-
+      method: 'POST', headers: getLiteLLMHeaders(), body: form
-      if (isTranscriptionModel) {
+    }).then(function(r) { return r.json().then(function(d) { return { data: d }; }); });
-        var ext = mime.split('/')[1] || 'webm';
+    text = sttResp.data && sttResp.data.text ? sttResp.data.text : '';
        var file = new File([audioBuffer], 'audio.' + ext, { type: mime });
        var form = new FormData();
        form.append('file', file);
        form.append('model', sttModel);
        var sttResp = await fetch(gatewayUrl('/audio/transcriptions'), {
          method: 'POST', headers: getLiteLLMHeaders(), body: form
        }).then(function(r) { return r.json().then(function(d) { return { data: d }; }); });
        text = sttResp.data && sttResp.data.text ? sttResp.data.text : '';
      } else {
        var sttResp = await axios.post(gatewayUrl('/chat/completions'), {
          model: sttModel,
          messages: [{ role: 'user', content: [
            { type: 'input_audio', input_audio: { data: audioBase64, format: mime.split('/')[1] || 'webm' } },
            { type: 'text', text: 'Transcribe this audio. Output the spoken words only.' }
          ]}]
        }, { headers: getLiteLLMHeaders('application/json'), timeout: 60000 });
        if (sttResp.data && sttResp.data.choices && sttResp.data.choices[0]) {
          text = (sttResp.data.choices[0].message && sttResp.data.choices[0].message.content) || '';
        }
      }
    } else if (provider === 'openai') {
      if (!whisperClient) return res.json({ success: false, error: 'Whisper not configured' });
      var file = new File([audioBuffer], 'audio.webm', { type: mime });
      var wResult = await whisperClient.audio.transcriptions.create({
        file: file, model: 'whisper-1', language: 'en', response_format: 'text'
      });
      text = typeof wResult === 'string' ? wResult : (wResult.text || '');
    }
    res.json({ success: true, text: text.trim(), provider: provider, duration: Date.now() - start });
  } catch (e) {
@ -794,8 +652,6 @@ router.get('/config/embeddings', async function(req, res) {
    var provider = 'none';
    if (process.env.LITELLM_API_BASE) provider = 'litellm';
    else if (process.env.GOOGLE_APPLICATION_CREDENTIALS || process.env.VERTEX_PROJECT || process.env.GOOGLE_VERTEX_PROJECT) provider = 'vertex';
    else if (process.env.OPENAI_API_KEY) provider = 'openai';
    res.json({
      success: true,
@ -807,14 +663,7 @@ router.get('/config/embeddings', async function(req, res) {
      dbDimensions: dbDims,
      envModel: envModel,
      envDimensions: envDims,
-      models: [
+      models: []
        { id: 'vertex_ai/text-embedding-005', name: 'Vertex text-embedding-005', dims: 768, tag: 'RECOMMENDED' },
        { id: 'vertex_ai/gemini-embedding-001', name: 'Vertex gemini-embedding-001', dims: 3072, tag: 'HIGH-DIM' },
        { id: 'vertex_ai/text-multilingual-embedding-002', name: 'Vertex multilingual-embedding-002', dims: 768, tag: 'MULTILINGUAL' },
        { id: 'text-embedding-3-small', name: 'OpenAI text-embedding-3-small', dims: 1536, tag: 'OPENAI' },
        { id: 'text-embedding-3-large', name: 'OpenAI text-embedding-3-large', dims: 3072, tag: 'OPENAI' },
        { id: 'text-embedding-ada-002', name: 'OpenAI text-embedding-ada-002', dims: 1536, tag: 'OPENAI' }
      ]
    });
  } catch (e) { res.status(500).json({ error: 'Request failed' }); }
 });
@ -828,15 +677,6 @@ router.get('/config/embeddings/discover', async function(req, res) {
    var provider = 'none';
    if (process.env.LITELLM_API_BASE) provider = 'litellm';
    else if (process.env.GOOGLE_APPLICATION_CREDENTIALS || process.env.VERTEX_PROJECT || process.env.GOOGLE_VERTEX_PROJECT) provider = 'vertex';
    else if (process.env.OPENAI_API_KEY) provider = 'openai';
    var knownVertex = [
      { id: 'vertex_ai/text-embedding-005', name: 'Vertex text-embedding-005', dims: 768 },
      { id: 'vertex_ai/gemini-embedding-001', name: 'Vertex gemini-embedding-001', dims: 3072 },
      { id: 'vertex_ai/text-multilingual-embedding-002', name: 'Vertex multilingual-embedding-002', dims: 768 },
      { id: 'text-embedding-005', name: 'text-embedding-005 (direct)', dims: 768 }
    ];
    if (provider === 'litellm') {
      try {
@ -846,37 +686,6 @@ router.get('/config/embeddings/discover', async function(req, res) {
          discovered.push({ id: m.id, name: m.name, dims: m.dims, source: 'gateway-api' });
        });
      } catch(e) { logger.warn('LiteLLM embedding metadata discovery failed: ' + e.message); }
      if (discovered.length === 0) {
        knownVertex.forEach(function(m) {
          if (!discovered.find(function(d) { return d.id === m.id; })) {
            discovered.push(Object.assign({ source: 'builtin' }, m));
          }
        });
      }
    }
    if (provider === 'vertex') {
      knownVertex.forEach(function(m) {
        discovered.push(Object.assign({ source: 'vertex-builtin' }, m));
      });
    }
    if (provider === 'openai') {
      try {
        var OpenAI = require('openai').OpenAI;
        var oc = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
        var oMods = await oc.models.list();
        oMods.data.forEach(function(m) {
          if (m.id.toLowerCase().indexOf('embedding') !== -1) {
            var dims = m.id === 'text-embedding-3-large' ? 3072 : m.id === 'text-embedding-3-small' ? 1536 : 1536;
            discovered.push({ id: m.id, name: m.id, dims: dims, source: 'openai-api' });
          }
        });
      } catch(e) {
        [{ id: 'text-embedding-3-small', dims: 1536 }, { id: 'text-embedding-3-large', dims: 3072 }, { id: 'text-embedding-ada-002', dims: 1536 }].forEach(function(m) {
          discovered.push({ id: m.id, name: m.id, dims: m.dims, source: 'openai-builtin' });
        });
      }
    }
    if (search) {
--- a/src/routes/transcribe.js
+++ b/src/routes/transcribe.js
@ -1,52 +1,30 @@
 const express = require('express');
 const router = express.Router();
 const multer = require('multer');
-const { whisperClient, litellmClient } = require('../utils/ai');
+const { litellmClient } = require('../utils/ai');
 const { transcribeWithAWS, isAWSTranscribeConfigured } = require('../utils/transcribeAWS');
 const { transcribeWithLocal, isLocalWhisperConfigured } = require('../utils/transcribeLocal');
 const { transcribeWithGemini, isGoogleSTTConfigured } = require('../utils/transcribeGoogle');
 const { authMiddleware } = require('../middleware/auth');
 var logger = require('../utils/logger');
 var { gatewayUrl } = require('../utils/errors');
 var { getLiteLLMHeaders } = require('../utils/litellm');
 const upload = multer({ storage: multer.memoryStorage(), limits: { fileSize: 25 * 1024 * 1024 } });
-// Provider priority (auto-detect):
+// STT is intentionally routed only through LiteLLM.
 // TRANSCRIBE_PROVIDER=google  → Vertex AI / Gemini (direct, HIPAA eligible)
 // TRANSCRIBE_PROVIDER=aws     → Amazon Transcribe (HIPAA eligible)
 // TRANSCRIBE_PROVIDER=local   → local whisper.cpp / faster-whisper
 // TRANSCRIBE_PROVIDER=openai  → OpenAI Whisper (direct)
 // TRANSCRIBE_PROVIDER=litellm → LiteLLM proxy at /v1/audio/transcriptions
 // Auto: google > aws > openai
 function getTranscribeProvider() {
  var env = process.env.TRANSCRIBE_PROVIDER;
  if (env === 'google')  return 'google';
  if (env === 'aws')     return 'aws';
  if (env === 'local')   return 'local';
  if (env === 'openai')  return 'openai';
  if (env === 'litellm') return 'litellm';
-  // Auto-detect
+  if (litellmClient) return 'litellm';
-  if (isGoogleSTTConfigured())    return 'google';
+  return 'none';
  if (isAWSTranscribeConfigured()) return 'aws';
  return 'openai';
 }
 function isTranscribeAvailable() {
  if (isGoogleSTTConfigured())    return true;
  if (isLocalWhisperConfigured()) return true;
  if (isAWSTranscribeConfigured()) return true;
  if (litellmClient) return true;
  if (whisperClient) return true;
  return false;
 }
 var provider = getTranscribeProvider();
 var medical  = process.env.AWS_TRANSCRIBE_MEDICAL === 'true';
 var available = isTranscribeAvailable();
-console.log('🎙️  Transcribe provider:', provider +
+console.log('🎙️  Transcribe provider:', provider + (available ? '' : ' (NOT CONFIGURED — browser speech only)'));
  (provider === 'aws' && medical ? ' (Medical)' : '') +
  (provider === 'google' ? ' (model: ' + (process.env.GOOGLE_STT_MODEL || 'gemini-2.0-flash') + ')' : '') +
  (available ? '' : ' (NOT CONFIGURED — browser speech only)'));
 router.get('/transcribe/status', authMiddleware, (req, res) => {
  res.json({ available: available, provider: available ? provider : 'none' });
@ -66,68 +44,29 @@ router.post('/transcribe', authMiddleware, upload.single('audio'), async (req, r
    console.log('[Transcribe] Received ' + (fileSize / 1024).toFixed(0) + 'KB audio (' + (req.file.mimetype || 'unknown') + ') via ' + provider + (userModel ? ' (user model: ' + userModel + ')' : ''));
-    if (provider === 'google') {
+    if (provider !== 'litellm' || !process.env.LITELLM_API_BASE) return res.status(400).json({ error: 'LITELLM_API_BASE not set.' });
-      var model = userModel || adminSttModel || process.env.GOOGLE_STT_MODEL || 'gemini-2.0-flash';
+    var sttModel = userModel || adminSttModel || process.env.LITELLM_STT_MODEL || '';
-      var text = await transcribeWithGemini(req.file.buffer, req.file.mimetype || 'audio/webm', model);
+    if (!sttModel) return res.status(400).json({ error: 'No LiteLLM STT model configured.' });
-      console.log('[Transcribe] Google/' + model + ' done in ' + (Date.now() - startTime) + 'ms');
+    var mimeType = req.file.mimetype || 'audio/webm';
-      logger.audit(req.user.id, 'transcribe', 'Transcribed audio via google', req, { category: 'clinical' });
+    var ext = mimeType.split('/')[1] || 'webm';
      return res.json({ success: true, text: text, provider: 'google-' + model, duration: Date.now() - startTime });
    }
-    if (provider === 'local') {
+    var file = new File([req.file.buffer], 'audio.' + ext, { type: mimeType });
-      var text = await transcribeWithLocal(req.file.buffer, req.file.mimetype || 'audio/webm');
+    var form = new FormData();
-      console.log('[Transcribe] Local done in ' + (Date.now() - startTime) + 'ms');
+    form.append('file', file);
-      logger.audit(req.user.id, 'transcribe', 'Transcribed audio via local', req, { category: 'clinical' });
+    form.append('model', sttModel);
      return res.json({ success: true, text: text, provider: 'local-whisper', duration: Date.now() - startTime });
    }
-    if (provider === 'aws') {
+    var sttResp = await fetch(gatewayUrl('/audio/transcriptions'), {
-      if (!isAWSTranscribeConfigured()) return res.status(400).json({ error: 'AWS Transcribe not configured.' });
+      method: 'POST', headers: getLiteLLMHeaders(), body: form,
      var text = await transcribeWithAWS(req.file.buffer, req.file.mimetype || 'audio/webm');
      console.log('[Transcribe] AWS done in ' + (Date.now() - startTime) + 'ms');
      logger.audit(req.user.id, 'transcribe', 'Transcribed audio via aws', req, { category: 'clinical' });
      return res.json({ success: true, text: text, provider: 'aws-transcribe', duration: Date.now() - startTime });
    }
    if (provider === 'litellm') {
      if (!process.env.LITELLM_API_BASE) return res.status(400).json({ error: 'LITELLM_API_BASE not set.' });
      var sttModel = userModel || adminSttModel || process.env.LITELLM_STT_MODEL || 'whisper-1';
      var mimeType = req.file.mimetype || 'audio/webm';
      var ext = mimeType.split('/')[1] || 'webm';
      var file = new File([req.file.buffer], 'audio.' + ext, { type: mimeType });
      var form = new FormData();
      form.append('file', file);
      form.append('model', sttModel);
      var fetchHeaders = {};
      if (process.env.LITELLM_API_KEY) fetchHeaders['Authorization'] = 'Bearer ' + process.env.LITELLM_API_KEY;
      var sttResp = await fetch(gatewayUrl('/audio/transcriptions'), {
        method: 'POST', headers: fetchHeaders, body: form,
      });
      if (!sttResp.ok) {
        var errBody = await sttResp.text();
        throw new Error('LiteLLM /audio/transcriptions ' + sttResp.status + ': ' + errBody.substring(0, 500));
      }
      var data = await sttResp.json();
      var text = (data && data.text) ? String(data.text).trim() : '';
      console.log('[Transcribe] LiteLLM/' + sttModel + ' done in ' + (Date.now() - startTime) + 'ms');
      logger.audit(req.user.id, 'transcribe', 'Transcribed audio via litellm', req, { category: 'clinical' });
      return res.json({ success: true, text: text, provider: 'litellm/' + sttModel, duration: Date.now() - startTime });
    }
    // OpenAI Whisper (direct)
    if (!whisperClient) return res.status(400).json({ error: 'Whisper not configured. Set OPENAI_API_KEY.' });
    var file   = new File([req.file.buffer], 'audio.webm', { type: req.file.mimetype || 'audio/webm' });
    var result = await whisperClient.audio.transcriptions.create({
      file, model: 'whisper-1', language: 'en',
      response_format: 'text',
      prompt: 'Medical patient encounter. Pediatric. Clinical terms, diagnoses, medications.'
    });
-    var text = typeof result === 'string' ? result : result.text;
+    if (!sttResp.ok) {
-    console.log('[Transcribe] Whisper done in ' + (Date.now() - startTime) + 'ms');
+      var errBody = await sttResp.text();
-    logger.audit(req.user.id, 'transcribe', 'Transcribed audio via openai', req, { category: 'clinical' });
+      throw new Error('LiteLLM /audio/transcriptions ' + sttResp.status + ': ' + errBody.substring(0, 500));
-    res.json({ success: true, text: text, provider: 'openai-whisper', duration: Date.now() - startTime });
+    }
    var data = await sttResp.json();
    var text = (data && data.text) ? String(data.text).trim() : '';
    console.log('[Transcribe] LiteLLM/' + sttModel + ' done in ' + (Date.now() - startTime) + 'ms');
    logger.audit(req.user.id, 'transcribe', 'Transcribed audio via litellm', req, { category: 'clinical' });
    return res.json({ success: true, text: text, provider: 'litellm/' + sttModel, duration: Date.now() - startTime });
  } catch (err) {
    var detail = err.response && err.response.data
--- a/src/routes/tts.js
+++ b/src/routes/tts.js
@ -1,26 +1,15 @@
 const express = require('express');
 const router  = express.Router();
 const axios   = require('axios');
 const { synthesizeWithGoogleTTS } = require('../utils/ttsGoogle');
 const { authMiddleware } = require('../middleware/auth');
 var logger = require('../utils/logger');
 var { gatewayUrl } = require('../utils/errors');
 var { getTTSProvider } = require('../utils/ttsProvider');
 var { getLiteLLMHeaders } = require('../utils/litellm');
-// Provider priority (auto-detect):
+// TTS is intentionally routed only through LiteLLM. Provider-specific voice
-// TTS_PROVIDER=litellm    → LiteLLM proxy at /v1/audio/speech
+// routing belongs in LiteLLM config, not this app.
 // TTS_PROVIDER=google     → Google Cloud TTS direct (HIPAA eligible, opt-in)
 // TTS_PROVIDER=elevenlabs → ElevenLabs direct (not HIPAA, opt-in)
 // Auto: litellm > google > elevenlabs
 //
 // Design: the LiteLLM branch is a pure passthrough — model + voice strings
 // go straight into the OpenAI-compatible /v1/audio/speech body. No regex
 // dispatch on model/voice names; the LiteLLM config decides which provider
 // (OpenAI, ElevenLabs, Vertex, Azure) actually serves the request.
 var ttsProvider = getTTSProvider();
-console.log('🔊 TTS provider:', ttsProvider +
+console.log('🔊 TTS provider:', ttsProvider);
  (ttsProvider === 'google' ? ' (voice: ' + (process.env.GOOGLE_TTS_VOICE || 'en-US-Journey-F') + ')' : ''));
 router.post('/text-to-speech', authMiddleware, async (req, res) => {
  try {
@ -35,51 +24,27 @@ router.post('/text-to-speech', authMiddleware, async (req, res) => {
    var adminModel = await db.getSetting('tts.model') || '';
    var resolvedVoice = userVoice || adminVoice || '';
-    if (ttsProvider === 'litellm') {
+    if (ttsProvider !== 'litellm' || !process.env.LITELLM_API_BASE) {
-      if (!process.env.LITELLM_API_BASE) return res.status(400).json({ error: 'LITELLM_API_BASE not set.' });
+      return res.status(400).json({ error: 'TTS not configured. Set LITELLM_API_BASE.' });
      var ttsModel = adminModel || process.env.LITELLM_TTS_MODEL || 'tts-1';
      var ttsVoice = resolvedVoice || process.env.LITELLM_TTS_VOICE || 'alloy';
      var ttsResp = await fetch(gatewayUrl('/audio/speech'), {
        method: 'POST',
        headers: getLiteLLMHeaders('application/json'),
        body: JSON.stringify({ model: ttsModel, input: text, voice: ttsVoice })
      });
      if (!ttsResp.ok) {
        var errBody = await ttsResp.text();
        throw new Error('LiteLLM /audio/speech ' + ttsResp.status + ': ' + errBody.substring(0, 500));
      }
      var audioBuf = Buffer.from(await ttsResp.arrayBuffer());
      res.set('Content-Type', ttsResp.headers.get('content-type') || 'audio/mpeg');
      res.set('X-TTS-Provider', 'litellm/' + ttsModel);
      logger.audit(req.user.id, 'text_to_speech', 'TTS generated', req, { category: 'clinical' });
      return res.send(audioBuf);
    }
    var ttsModel = adminModel || process.env.LITELLM_TTS_MODEL || '';
    var ttsVoice = resolvedVoice || process.env.LITELLM_TTS_VOICE || '';
    if (!ttsModel) return res.status(400).json({ error: 'No LiteLLM TTS model configured.' });
-    if (ttsProvider === 'google') {
+    var ttsResp = await fetch(gatewayUrl('/audio/speech'), {
-      var voice = resolvedVoice || process.env.GOOGLE_TTS_VOICE || 'en-US-Journey-F';
+      method: 'POST',
-      var buffer = await synthesizeWithGoogleTTS(text, voice);
+      headers: getLiteLLMHeaders('application/json'),
-      res.set('Content-Type', 'audio/mpeg');
+      body: JSON.stringify({ model: ttsModel, input: text, voice: ttsVoice })
-      res.set('X-TTS-Provider', 'google-tts/' + voice);
+    });
-      logger.audit(req.user.id, 'text_to_speech', 'TTS generated', req, { category: 'clinical' });
+    if (!ttsResp.ok) {
-      return res.send(buffer);
+      var errBody = await ttsResp.text();
      throw new Error('LiteLLM /audio/speech ' + ttsResp.status + ': ' + errBody.substring(0, 500));
    }
-
+    var audioBuf = Buffer.from(await ttsResp.arrayBuffer());
-    if (ttsProvider === 'elevenlabs') {
+    res.set('Content-Type', ttsResp.headers.get('content-type') || 'audio/mpeg');
-      if (!process.env.ELEVENLABS_API_KEY) return res.status(400).json({ error: 'ElevenLabs not configured' });
+    res.set('X-TTS-Provider', 'litellm/' + ttsModel);
-      var resp = await axios({
+    logger.audit(req.user.id, 'text_to_speech', 'TTS generated', req, { category: 'clinical' });
-        method: 'POST',
+    return res.send(audioBuf);
        url: 'https://api.elevenlabs.io/v1/text-to-speech/pNInz6obpgDQGcFmaJgB',
        headers: { 'xi-api-key': process.env.ELEVENLABS_API_KEY, 'Content-Type': 'application/json' },
        data: { text: text, model_id: 'eleven_turbo_v2_5', voice_settings: { stability: 0.5, similarity_boost: 0.75 } },
        responseType: 'arraybuffer'
      });
      res.set('Content-Type', 'audio/mpeg');
      res.set('X-TTS-Provider', 'elevenlabs');
      logger.audit(req.user.id, 'text_to_speech', 'TTS generated', req, { category: 'clinical' });
      return res.send(Buffer.from(resp.data));
    }
    res.status(400).json({ error: 'TTS not configured. Set LITELLM_API_BASE, GOOGLE_VERTEX_PROJECT, or ELEVENLABS_API_KEY.' });
  } catch (err) {
    var detail = err.response && err.response.data
      ? JSON.stringify(err.response.data).substring(0, 500)
--- a/src/routes/userPreferences.js
+++ b/src/routes/userPreferences.js
@ -6,6 +6,8 @@ var express = require('express');
 var router = express.Router();
 var db = require('../db/database');
 var { authMiddleware } = require('../middleware/auth');
 var { getSTTModelLists, getSTTProvider } = require('../utils/sttProvider');
 var { getTTSProvider, getTTSVoiceLists } = require('../utils/ttsProvider');
 router.use(authMiddleware);
@ -44,103 +46,11 @@ router.post('/preferences', async function(req, res) {
 // Get available STT models and TTS voices
 router.get('/preferences/options', async function(req, res) {
  try {
-    var provider = process.env.TRANSCRIBE_PROVIDER;
+    var provider = getSTTProvider();
-    var ttsProvider = process.env.TTS_PROVIDER;
+    var ttsProvider = getTTSProvider();
-    // Auto-detect providers if not explicitly set
+    var sttModels = getSTTModelLists().litellm.map(function(model) { return { value: model, label: model }; });
-    if (!provider) {
+    var ttsVoices = getTTSVoiceLists().litellm.map(function(voice) { return { value: voice, label: voice }; });
      if (process.env.VERTEX_PROJECT || process.env.GOOGLE_CLOUD_PROJECT) provider = 'google';
      else if (process.env.AWS_BEDROCK_REGION) provider = 'aws';
      else if (process.env.LITELLM_API_BASE) provider = 'litellm';
      else if (process.env.OPENAI_API_KEY) provider = 'openai';
    }
    if (!ttsProvider) {
      if (process.env.VERTEX_PROJECT || process.env.GOOGLE_CLOUD_PROJECT) ttsProvider = 'google';
      else if (process.env.LITELLM_API_BASE) ttsProvider = 'litellm';
      else if (process.env.OPENAI_API_KEY) ttsProvider = 'openai';
      else if (process.env.ELEVENLABS_API_KEY) ttsProvider = 'elevenlabs';
    }
    // STT Models
    var sttModels = [];
    if (provider === 'google') {
      sttModels = [
        { value: 'gemini-2.0-flash-exp', label: 'Gemini 2.0 Flash (Experimental, fastest)' },
        { value: 'gemini-2.0-flash', label: 'Gemini 2.0 Flash (Fast, accurate)' },
        { value: 'gemini-1.5-flash', label: 'Gemini 1.5 Flash (Stable)' },
        { value: 'gemini-1.5-pro', label: 'Gemini 1.5 Pro (Best quality, slower)' }
      ];
    } else if (provider === 'litellm') {
      sttModels = [
        { value: 'local-whisper-large-v3-turbo', label: 'Local Whisper Large v3 Turbo' },
        { value: 'local-whisper-large-v3', label: 'Local Whisper Large v3' },
        { value: 'local-parakeet-v3', label: 'Local Parakeet v3' },
        { value: 'local-whisper-1', label: 'Local Whisper-1 Alias' },
        { value: 'gemini-2.0-flash-exp', label: 'Gemini 2.0 Flash Exp' },
        { value: 'gemini-2.0-flash', label: 'Gemini 2.0 Flash' },
        { value: 'gemini-1.5-flash', label: 'Gemini 1.5 Flash' },
        { value: 'gemini-1.5-pro', label: 'Gemini 1.5 Pro' }
      ];
    } else if (provider === 'openai') {
      sttModels = [
        { value: 'whisper-1', label: 'Whisper-1 (OpenAI standard)' }
      ];
    } else if (provider === 'aws') {
      sttModels = [
        { value: 'default', label: 'AWS Transcribe (Standard)' },
        { value: 'medical', label: 'AWS Transcribe Medical' }
      ];
    } else if (provider === 'local') {
      sttModels = [
        { value: 'tiny', label: 'Whisper Tiny (fastest)' },
        { value: 'base', label: 'Whisper Base' },
        { value: 'small', label: 'Whisper Small' },
        { value: 'medium', label: 'Whisper Medium' },
        { value: 'large', label: 'Whisper Large (best quality)' }
      ];
    }
    // TTS Voices
    var ttsVoices = [];
    if (ttsProvider === 'google') {
      ttsVoices = [
        { value: 'en-US-Journey-F', label: 'Journey (Female, natural)' },
        { value: 'en-US-Journey-D', label: 'Journey (Male, natural)' },
        { value: 'en-US-Studio-O', label: 'Studio O (Female, expressive)' },
        { value: 'en-US-Studio-M', label: 'Studio M (Male, expressive)' },
        { value: 'en-US-Neural2-A', label: 'Neural2 A (Male, standard)' },
        { value: 'en-US-Neural2-C', label: 'Neural2 C (Female, standard)' },
        { value: 'en-US-Neural2-D', label: 'Neural2 D (Male, standard)' },
        { value: 'en-US-Neural2-E', label: 'Neural2 E (Female, standard)' },
        { value: 'en-US-Neural2-F', label: 'Neural2 F (Female, standard)' },
        { value: 'en-US-Neural2-G', label: 'Neural2 G (Female, standard)' },
        { value: 'en-US-Neural2-H', label: 'Neural2 H (Female, standard)' },
        { value: 'en-US-Neural2-I', label: 'Neural2 I (Male, standard)' },
        { value: 'en-US-Neural2-J', label: 'Neural2 J (Male, standard)' }
      ];
    } else if (ttsProvider === 'litellm' || ttsProvider === 'openai') {
      ttsVoices = [
        { value: 'alloy', label: 'Alloy (Neutral)' },
        { value: 'echo', label: 'Echo (Male)' },
        { value: 'fable', label: 'Fable (British Male)' },
        { value: 'onyx', label: 'Onyx (Deep Male)' },
        { value: 'nova', label: 'Nova (Female)' },
        { value: 'shimmer', label: 'Shimmer (Soft Female)' }
      ];
    } else if (ttsProvider === 'elevenlabs') {
      ttsVoices = [
        { value: 'adam', label: 'Adam (Male, deep)' },
        { value: 'rachel', label: 'Rachel (Female, calm)' },
        { value: 'domi', label: 'Domi (Female, strong)' },
        { value: 'bella', label: 'Bella (Female, soft)' },
        { value: 'antoni', label: 'Antoni (Male, deep)' },
        { value: 'elli', label: 'Elli (Female, young)' },
        { value: 'josh', label: 'Josh (Male, narration)' },
        { value: 'arnold', label: 'Arnold (Male, crisp)' },
        { value: 'sam', label: 'Sam (Male, raspy)' }
      ];
    }
    res.json({
      success: true,
--- a/src/utils/ai.js
+++ b/src/utils/ai.js
@ -110,17 +110,6 @@ if (process.env.LITELLM_API_BASE) {
  }
 }
 // ============================================================
 // WHISPER CLIENT (always OpenAI, separate from text AI)
 // ============================================================
 var whisperClient = null;
 if (process.env.OPENAI_API_KEY && process.env.OPENAI_API_KEY.startsWith('sk-')) {
  whisperClient = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
  console.log('✅ Whisper: configured');
 } else {
  console.log('⚠️  Whisper: not configured (set OPENAI_API_KEY)');
 }
 // Force provider from env if explicitly set
 if (process.env.AI_PROVIDER) {
  activeProvider = process.env.AI_PROVIDER;
@ -694,4 +683,4 @@ async function discoverModels() {
  return discovered;
 }
-module.exports = { callAI, callAIStream, whisperClient, activeProvider, discoverModels, vertexClient, litellmClient };
+module.exports = { callAI, callAIStream, activeProvider, discoverModels, vertexClient, litellmClient };
--- a/src/utils/embeddings.js
+++ b/src/utils/embeddings.js
@ -1,18 +1,13 @@
 // ============================================================
-// EMBEDDINGS UTILITY — Generate & search with Vertex AI embeddings
+// EMBEDDINGS UTILITY — Generate & search through LiteLLM embeddings
 // Supports: Vertex AI (direct), LiteLLM proxy, OpenAI fallback
 // ============================================================
 var axios = require('axios');
 var { gatewayUrl } = require('./errors');
 var { getLiteLLMHeaders } = require('./litellm');
-// Vertex AI embedding models (via LiteLLM or direct)
+var DEFAULT_MODEL = 'openai-text-embedding-3-large';
-// gemini-embedding-001: 768 dims, multilingual + code, best quality
+var DEFAULT_DIMS = 3072;
 // text-embedding-005: 768 dims, English + code optimized
 // text-multilingual-embedding-002: 768 dims, multilingual focus
 var DEFAULT_MODEL = 'vertex_ai/text-embedding-005';
 var DEFAULT_DIMS = 768;
 /**
 * Generate embedding for text using configured provider
@ -40,22 +35,11 @@ async function generateEmbedding(text, opts) {
    throw new Error('Empty text provided for embedding');
  }
  // Try LiteLLM first if configured
  if (process.env.LITELLM_API_BASE) {
    return await generateEmbeddingLiteLLM(truncated, model, dimensions);
  }
-  // Try Vertex AI direct if configured
+  throw new Error('No embedding provider configured. Set LITELLM_API_BASE');
  if (process.env.GOOGLE_APPLICATION_CREDENTIALS || process.env.VERTEX_PROJECT) {
    return await generateEmbeddingVertexDirect(truncated, model, dimensions);
  }
  // Fallback to OpenAI if configured
  if (process.env.OPENAI_API_KEY) {
    return await generateEmbeddingOpenAI(truncated, model, dimensions);
  }
  throw new Error('No embedding provider configured. Set LITELLM_API_BASE, VERTEX_PROJECT, or OPENAI_API_KEY');
 }
 /**
@ -68,8 +52,7 @@ async function generateEmbeddingLiteLLM(text, model, dimensions) {
      input: text
    };
-    // Only include dimensions if model supports it (some models have fixed dims)
+    if (dimensions) {
    if (dimensions && model.includes('text-embedding-005')) {
      payload.dimensions = dimensions;
    }
@ -89,80 +72,6 @@ async function generateEmbeddingLiteLLM(text, model, dimensions) {
  }
 }
 /**
 * Generate embedding via Vertex AI direct (using @google-cloud/vertexai)
 */
 async function generateEmbeddingVertexDirect(text, model, dimensions) {
  try {
    var { VertexAI } = require('@google-cloud/vertexai');
    var project = process.env.VERTEX_PROJECT || process.env.GOOGLE_CLOUD_PROJECT;
    var location = process.env.VERTEX_LOCATION || 'us-central1';
    if (!project) {
      throw new Error('VERTEX_PROJECT or GOOGLE_CLOUD_PROJECT not set');
    }
    var vertexAI = new VertexAI({ project: project, location: location });
    // Extract model name (strip vertex_ai/ prefix if present)
    var modelName = model.replace(/^vertex_ai\//, '');
    // For text-embedding-005, we can specify output dimensions
    var request = {
      instances: [{ content: text }]
    };
    if (dimensions && modelName.includes('text-embedding-005')) {
      request.parameters = { outputDimensionality: dimensions };
    }
    // Use predictText API for embeddings
    var predictionClient = vertexAI.preview.getPredictionServiceClient();
    var endpoint = `projects/${project}/locations/${location}/publishers/google/models/${modelName}`;
    var [response] = await predictionClient.predict({
      endpoint: endpoint,
      instances: [{ content: text }],
      parameters: request.parameters || {}
    });
    if (!response || !response.predictions || !response.predictions[0]) {
      throw new Error('Invalid response from Vertex AI');
    }
    var prediction = response.predictions[0];
    return prediction.embeddings?.values || prediction.values || prediction;
  } catch (err) {
    console.error('[Embeddings] Vertex AI direct error:', err.message);
    throw new Error('Vertex AI embedding failed: ' + err.message);
  }
 }
 /**
 * Generate embedding via OpenAI (fallback)
 */
 async function generateEmbeddingOpenAI(text, model, dimensions) {
  try {
    var openai = require('openai');
    var client = new openai.OpenAI({ apiKey: process.env.OPENAI_API_KEY });
    // Use OpenAI's text-embedding-3-small model (1536 dims by default)
    var embModel = 'text-embedding-3-small';
    var response = await client.embeddings.create({
      model: embModel,
      input: text,
      dimensions: dimensions || 768 // OpenAI supports custom dimensions
    });
    return response.data[0].embedding;
  } catch (err) {
    console.error('[Embeddings] OpenAI error:', err.message);
    throw new Error('OpenAI embedding failed: ' + err.message);
  }
 }
 /**
 * Search for similar content using cosine similarity
 * @param {string} queryText - Search query
@ -271,12 +180,7 @@ function getLiteLLMEmbeddingModels(models) {
 * Check if embeddings are available (provider configured)
 */
 function isEmbeddingsAvailable() {
-  return !!(
+  return !!process.env.LITELLM_API_BASE;
    process.env.LITELLM_API_BASE ||
    process.env.VERTEX_PROJECT ||
    process.env.GOOGLE_CLOUD_PROJECT ||
    process.env.OPENAI_API_KEY
  );
 }
 module.exports = {
--- a/src/utils/sttProvider.js
+++ b/src/utils/sttProvider.js
@ -1,43 +1,24 @@
 const GOOGLE_STT_MODELS = ['gemini-2.0-flash', 'gemini-2.5-flash', 'gemini-2.0-flash-lite'];
 const AWS_STT_MODES = ['standard', 'medical'];
 const OPENAI_STT_MODELS = ['whisper-1'];
 const LITELLM_STT_MODELS = ['local-whisper-large-v3-turbo', 'local-whisper-large-v3', 'local-parakeet-v3', 'local-whisper-1', 'gemini-2.0-flash', 'gemini-2.5-flash'];
 const LOCAL_WHISPER_MODELS = ['tiny', 'base', 'small', 'medium', 'large'];
 function getSTTDependencies() {
-  var { isGoogleSTTConfigured } = require('./transcribeGoogle');
+  var { litellmClient } = require('./ai');
  var { isAWSTranscribeConfigured } = require('./transcribeAWS');
  var { isLocalWhisperConfigured } = require('./transcribeLocal');
  var { whisperClient, litellmClient } = require('./ai');
  return {
-    google: isGoogleSTTConfigured(),
+    litellm: !!litellmClient
    aws: isAWSTranscribeConfigured(),
    local: isLocalWhisperConfigured(),
    litellm: !!litellmClient,
    openai: !!whisperClient
  };
 }
 function getSTTProvider(deps) {
  var env = process.env.TRANSCRIBE_PROVIDER || 'auto';
-  if (env !== 'auto') return env;
+  if (env === 'litellm') return 'litellm';
  var configured = deps || getSTTDependencies();
  if (configured.google) return 'google';
  if (configured.aws) return 'aws';
  if (configured.local) return 'local';
  if (configured.litellm) return 'litellm';
  if (configured.openai) return 'openai';
  return 'none';
 }
 function getSTTModelLists() {
  return {
-    google: GOOGLE_STT_MODELS,
+    litellm: LITELLM_STT_MODELS
    aws: AWS_STT_MODES,
    openai: OPENAI_STT_MODELS,
    litellm: LITELLM_STT_MODELS,
    local: LOCAL_WHISPER_MODELS
  };
 }
@ -53,11 +34,7 @@ function getLiteLLMSTTModels(models) {
 }
 module.exports = {
  AWS_STT_MODES,
  GOOGLE_STT_MODELS,
  LITELLM_STT_MODELS,
  LOCAL_WHISPER_MODELS,
  OPENAI_STT_MODELS,
  getSTTDependencies,
  getLiteLLMSTTModels,
  getSTTModelLists,
--- a/src/utils/transcribeAWS.js
+++ b/src/utils/transcribeAWS.js
@ -1,211 +0,0 @@
 // ============================================================
 // TRANSCRIBE-AWS.JS — Amazon Transcribe Streaming
 // No S3 required. Audio streams directly to AWS.
 //
 // Audio conversion: ffmpeg converts browser WebM/Opus → PCM 16kHz
 // which is the most reliable format for AWS Transcribe. If ffmpeg
 // is not installed, falls back to sending ogg-opus directly.
 //
 // Env vars:
 //   TRANSCRIBE_PROVIDER=aws          — use this instead of Whisper
 //   AWS_TRANSCRIBE_MEDICAL=true      — use Transcribe Medical (better
 //                                      accuracy for clinical dictation)
 //   AWS_TRANSCRIBE_SPECIALTY=PRIMARYCARE  — medical specialty
 //     Options: PRIMARYCARE, CARDIOLOGY, NEUROLOGY, ONCOLOGY,
 //              RADIOLOGY, UROLOGY
 //   AWS_BEDROCK_REGION               — reused for Transcribe region
 //   AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY — reused credentials
 // ============================================================
 const { spawn } = require('child_process');
 const CHUNK_SIZE = 8192; // 8 KB per audio chunk (AWS limit ~32KB per event frame)
 // Convert any browser audio (WebM, OGG, etc.) to raw PCM s16le 16kHz mono
 // using ffmpeg. Returns a Buffer of raw PCM bytes.
 // Throws if ffmpeg is not installed.
 function convertToPCM(inputBuffer) {
  return new Promise(function(resolve, reject) {
    var ff = spawn('ffmpeg', [
      '-i', 'pipe:0',       // read from stdin
      '-f', 's16le',        // raw signed 16-bit little-endian PCM
      '-ar', '16000',       // 16 kHz — ideal for speech recognition
      '-ac', '1',           // mono
      '-acodec', 'pcm_s16le',
      'pipe:1'              // write to stdout
    ]);
    var out = [];
    var errBuf = [];
    ff.stdout.on('data', function(d) { out.push(d); });
    ff.stderr.on('data', function(d) { errBuf.push(d); }); // ffmpeg logs to stderr, not an error
    ff.on('close', function(code) {
      if (code !== 0) {
        reject(new Error('ffmpeg conversion failed (code ' + code + '): ' + Buffer.concat(errBuf).toString().slice(-200)));
        return;
      }
      resolve(Buffer.concat(out));
    });
    ff.on('error', function(err) {
      reject(new Error('ffmpeg not found. Install ffmpeg on the server: ' + err.message));
    });
    ff.stdin.write(inputBuffer);
    ff.stdin.end();
  });
 }
 async function* makeAudioStream(buffer) {
  // 8KB chunks — larger sizes cause AWS SDK deserialization errors
  // ("to see the raw response, inspect the hidden field {error}.$response").
  // No artificial delay between chunks; yield with a microtask break
  // to avoid blocking the event loop without adding real latency.
  for (var i = 0; i < buffer.length; i += CHUNK_SIZE) {
    yield { AudioEvent: { AudioChunk: buffer.slice(i, i + CHUNK_SIZE) } };
    // Microtask break every 16 chunks (~128KB) to keep event loop responsive
    if ((i / CHUNK_SIZE) % 16 === 15) {
      await new Promise(function(r) { setImmediate(r); });
    }
  }
 }
 function logTranscribeError(label, err) {
  console.error('[Transcribe] ' + label + ':', err.message);
  if (err.name) console.error('[Transcribe]   name:', err.name);
  if (err.$metadata) console.error('[Transcribe]   metadata:', JSON.stringify(err.$metadata));
  if (err.cause) console.error('[Transcribe]   cause:', err.cause.message || JSON.stringify(err.cause));
  if (err.$response) {
    try {
      var resp = err.$response;
      console.error('[Transcribe]   status:', resp.statusCode);
      if (resp.headers) console.error('[Transcribe]   headers:', JSON.stringify(resp.headers));
      if (resp.body) console.error('[Transcribe]   body:', typeof resp.body === 'string' ? resp.body.slice(0, 500) : '(stream)');
    } catch (e) { /* ignore logging errors */ }
  }
 }
 async function transcribeWithAWS(audioBuffer, mimeType) {
  var startTime = Date.now();
  var TranscribeModule = require('@aws-sdk/client-transcribe-streaming');
  var TranscribeStreamingClient = TranscribeModule.TranscribeStreamingClient;
  var region = process.env.AWS_BEDROCK_REGION || process.env.AWS_REGION || 'us-east-1';
  var credentials = process.env.AWS_ACCESS_KEY_ID ? {
    accessKeyId: process.env.AWS_ACCESS_KEY_ID,
    secretAccessKey: process.env.AWS_SECRET_ACCESS_KEY
  } : undefined;
  // Try ffmpeg conversion to PCM (most reliable with AWS Transcribe).
  // Falls back to ogg-opus passthrough if ffmpeg is not installed.
  var audioToSend = audioBuffer;
  var mediaEncoding = 'ogg-opus';
  var sampleRate = 48000;
  try {
    var ffStart = Date.now();
    audioToSend = await convertToPCM(audioBuffer);
    mediaEncoding = 'pcm';
    sampleRate = 16000;
    console.log('[Transcribe] ffmpeg: ' + (Date.now() - ffStart) + 'ms — ' + audioBuffer.length + ' → ' + audioToSend.length + ' bytes');
  } catch (ffErr) {
    console.warn('[Transcribe] ffmpeg unavailable, sending ogg-opus directly:', ffErr.message);
    if (mimeType && mimeType.indexOf('wav') !== -1) {
      mediaEncoding = 'pcm';
      sampleRate = 16000;
    }
  }
  // Minimum ~0.5s of PCM audio at 16kHz mono 16-bit = 16000 bytes
  if (mediaEncoding === 'pcm' && audioToSend.length < 16000) {
    console.warn('[Transcribe] Audio too short (' + audioToSend.length + ' bytes), skipping');
    return '';
  }
  var useMedical = process.env.AWS_TRANSCRIBE_MEDICAL === 'true';
  var specialty = process.env.AWS_TRANSCRIBE_SPECIALTY || 'PRIMARYCARE';
  var transcript = '';
  // Helper to collect transcript from a result stream
  async function collectTranscript(resultStream) {
    var text = '';
    for await (var event of resultStream) {
      if (event.TranscriptEvent && event.TranscriptEvent.Transcript) {
        var results = event.TranscriptEvent.Transcript.Results || [];
        for (var i = 0; i < results.length; i++) {
          var r = results[i];
          if (!r.IsPartial && r.Alternatives && r.Alternatives[0]) {
            text += r.Alternatives[0].Transcript + ' ';
          }
        }
      }
    }
    return text;
  }
  // Try Medical first if enabled, with fallback to Standard
  var streamStart = Date.now();
  if (useMedical) {
    try {
      var client = new TranscribeStreamingClient({ region: region, credentials: credentials });
      var medCmd = new TranscribeModule.StartMedicalStreamTranscriptionCommand({
        LanguageCode: 'en-US',
        MediaSampleRateHertz: sampleRate,
        MediaEncoding: mediaEncoding,
        Specialty: specialty,
        Type: 'DICTATION',
        AudioStream: makeAudioStream(audioToSend)
      });
      var medResp = await client.send(medCmd);
      transcript = await collectTranscript(medResp.TranscriptResultStream);
      console.log('[Transcribe] Medical OK — ' + (Date.now() - streamStart) + 'ms, ' + transcript.length + ' chars');
    } catch (medErr) {
      logTranscribeError('Medical failed', medErr);
      console.warn('[Transcribe] Falling back to standard...');
      streamStart = Date.now();
      try {
        var client2 = new TranscribeStreamingClient({ region: region, credentials: credentials });
        var stdCmd = new TranscribeModule.StartStreamTranscriptionCommand({
          LanguageCode: 'en-US',
          MediaSampleRateHertz: sampleRate,
          MediaEncoding: mediaEncoding,
          AudioStream: makeAudioStream(audioToSend)
        });
        var stdResp = await client2.send(stdCmd);
        transcript = await collectTranscript(stdResp.TranscriptResultStream);
        console.log('[Transcribe] Standard fallback OK — ' + (Date.now() - streamStart) + 'ms, ' + transcript.length + ' chars');
      } catch (stdErr) {
        logTranscribeError('Standard also failed', stdErr);
        throw stdErr;
      }
    }
  } else {
    var client = new TranscribeStreamingClient({ region: region, credentials: credentials });
    var cmd = new TranscribeModule.StartStreamTranscriptionCommand({
      LanguageCode: 'en-US',
      MediaSampleRateHertz: sampleRate,
      MediaEncoding: mediaEncoding,
      AudioStream: makeAudioStream(audioToSend)
    });
    var resp = await client.send(cmd);
    transcript = await collectTranscript(resp.TranscriptResultStream);
    console.log('[Transcribe] Standard OK — ' + (Date.now() - streamStart) + 'ms, ' + transcript.length + ' chars');
  }
  console.log('[Transcribe] Total: ' + (Date.now() - startTime) + 'ms');
  return transcript.trim();
 }
 // Check if AWS Transcribe is available and configured
 function isAWSTranscribeConfigured() {
  try {
    require('@aws-sdk/client-transcribe-streaming');
    return !!(process.env.AWS_BEDROCK_REGION || process.env.AWS_REGION);
  } catch (e) {
    return false;
  }
 }
 module.exports = { transcribeWithAWS, isAWSTranscribeConfigured };
--- a/src/utils/transcribeGoogle.js
+++ b/src/utils/transcribeGoogle.js
@ -1,45 +0,0 @@
 // ============================================================
 // TRANSCRIBE GOOGLE — Vertex AI / Gemini audio transcription
 // Sends audio inline to Gemini which returns spoken text.
 // Supports: gemini-2.0-flash (default), gemini-2.5-flash
 // Requires: GOOGLE_VERTEX_PROJECT, @google-cloud/vertexai installed
 // ============================================================
 async function transcribeWithGemini(buffer, mimeType, modelName) {
  var VertexModule = require('@google-cloud/vertexai');
  var project  = process.env.GOOGLE_VERTEX_PROJECT;
  var location = process.env.GOOGLE_VERTEX_LOCATION || 'us-central1';
  if (!project) throw new Error('GOOGLE_VERTEX_PROJECT not set');
  var vertex = new VertexModule.VertexAI({ project: project, location: location });
  var model  = vertex.getGenerativeModel({
    model: modelName || process.env.GOOGLE_STT_MODEL || 'gemini-2.0-flash'
  });
  var base64   = buffer.toString('base64');
  var safeMime = mimeType || 'audio/webm';
  var result = await model.generateContent({
    contents: [{
      role: 'user',
      parts: [
        { inlineData: { mimeType: safeMime, data: base64 } },
        { text: 'Transcribe this audio. Output the spoken words only, exactly as heard. No commentary, no formatting, no explanation.' }
      ]
    }]
  });
  var text = '';
  if (result.response && result.response.candidates && result.response.candidates[0]) {
    var parts = result.response.candidates[0].content.parts || [];
    parts.forEach(function(p) { if (p.text) text += p.text; });
  }
  return text.trim();
 }
 function isGoogleSTTConfigured() {
  if (!process.env.GOOGLE_VERTEX_PROJECT) return false;
  try { require('@google-cloud/vertexai'); return true; } catch(e) { return false; }
 }
 module.exports = { transcribeWithGemini, isGoogleSTTConfigured };
--- a/src/utils/transcribeLocal.js
+++ b/src/utils/transcribeLocal.js
@ -1,178 +0,0 @@
 // ============================================================
 // LOCAL WHISPER TRANSCRIPTION — uses whisper.cpp or faster-whisper
 // ============================================================
 // Requires: whisper.cpp binary or faster-whisper-cli installed on the host.
 // Set TRANSCRIBE_PROVIDER=local and optionally WHISPER_MODEL_SIZE, WHISPER_BINARY.
 var { execFile } = require('child_process');
 var fs = require('fs');
 var path = require('path');
 var os = require('os');
 var crypto = require('crypto');
 var logger = require('./logger');
 var WHISPER_BINARY = process.env.WHISPER_BINARY || 'whisper-cpp';
 var WHISPER_MODEL_SIZE = process.env.WHISPER_MODEL_SIZE || 'small';
 var WHISPER_MODEL_PATH = process.env.WHISPER_MODEL_PATH || '';
 var WHISPER_LANGUAGE = process.env.WHISPER_LANGUAGE || 'en';
 var WHISPER_THREADS = process.env.WHISPER_THREADS || String(Math.max(2, os.cpus().length - 1));
 /**
 * Check if local Whisper is available
 */
 function isLocalWhisperConfigured() {
  return process.env.TRANSCRIBE_PROVIDER === 'local';
 }
 /**
 * Detect which whisper binary is available
 * Supports: whisper-cpp, whisper, faster-whisper
 */
 function detectBinary(callback) {
  var candidates = [WHISPER_BINARY, 'whisper-cpp', 'whisper', 'faster-whisper'];
  var tried = 0;
  function tryNext() {
    if (tried >= candidates.length) return callback(null);
    var bin = candidates[tried++];
    execFile(bin, ['--help'], { timeout: 5000 }, function(err) {
      if (!err || (err.code !== 'ENOENT' && err.code !== 127)) return callback(bin);
      tryNext();
    });
  }
  tryNext();
 }
 /**
 * Convert audio buffer to WAV using ffmpeg (whisper.cpp needs 16kHz mono WAV)
 */
 function convertToWav(inputPath, callback) {
  var wavPath = inputPath + '.wav';
  execFile('ffmpeg', [
    '-i', inputPath,
    '-ar', '16000',
    '-ac', '1',
    '-f', 'wav',
    '-y', wavPath
  ], { timeout: 30000 }, function(err) {
    if (err) return callback(err, null);
    callback(null, wavPath);
  });
 }
 /**
 * Transcribe audio buffer using local Whisper
 * @param {Buffer} audioBuffer - Audio data
 * @param {string} mimeType - MIME type of the audio
 * @returns {Promise<string>} Transcribed text
 */
 function transcribeWithLocal(audioBuffer, mimeType) {
  return new Promise(function(resolve, reject) {
    // Write buffer to temp file
    var ext = '.webm';
    if (mimeType && mimeType.includes('wav')) ext = '.wav';
    else if (mimeType && mimeType.includes('mp3')) ext = '.mp3';
    else if (mimeType && mimeType.includes('ogg')) ext = '.ogg';
    else if (mimeType && mimeType.includes('mp4')) ext = '.mp4';
    var tmpDir = os.tmpdir();
    var tmpFile = path.join(tmpDir, 'whisper_' + crypto.randomBytes(8).toString('hex') + ext);
    fs.writeFile(tmpFile, audioBuffer, function(err) {
      if (err) return reject(new Error('Failed to write temp audio: ' + err.message));
      // Convert to WAV for whisper.cpp compatibility
      convertToWav(tmpFile, function(convErr, wavPath) {
        var audioPath = convErr ? tmpFile : wavPath;
        detectBinary(function(binary) {
          if (!binary) {
            cleanup(tmpFile, wavPath);
            return reject(new Error(
              'Local Whisper not found. Install whisper.cpp (https://github.com/ggerganov/whisper.cpp) ' +
              'or faster-whisper (pip install faster-whisper). Set WHISPER_BINARY env var if using a custom path.'
            ));
          }
          var args = buildArgs(binary, audioPath);
          logger.info('[LocalWhisper] Running: ' + binary + ' ' + args.join(' '));
          execFile(binary, args, { timeout: 120000, maxBuffer: 10 * 1024 * 1024 }, function(execErr, stdout, stderr) {
            cleanup(tmpFile, wavPath);
            if (execErr) {
              logger.error('[LocalWhisper] Error:', execErr.message);
              if (stderr) logger.error('[LocalWhisper] stderr:', stderr.substring(0, 500));
              return reject(new Error('Local Whisper transcription failed: ' + execErr.message));
            }
            // Parse output — whisper.cpp outputs timestamped lines like [00:00:00.000 --> 00:00:05.000] text
            var text = parseOutput(stdout);
            if (!text || !text.trim()) {
              return reject(new Error('Local Whisper returned empty transcription'));
            }
            resolve(text.trim());
          });
        });
      });
    });
  });
 }
 /**
 * Build command-line args based on the detected binary
 */
 function buildArgs(binary, audioPath) {
  if (binary.includes('faster-whisper')) {
    // faster-whisper CLI
    var args = ['--model', WHISPER_MODEL_SIZE, '--language', WHISPER_LANGUAGE];
    if (WHISPER_THREADS) args.push('--threads', WHISPER_THREADS);
    args.push(audioPath);
    return args;
  }
  // whisper.cpp style
  var args = ['-f', audioPath, '-l', WHISPER_LANGUAGE, '-t', WHISPER_THREADS, '--no-timestamps'];
  if (WHISPER_MODEL_PATH) {
    args.push('-m', WHISPER_MODEL_PATH);
  } else {
    // whisper.cpp uses model names like ggml-small.bin
    args.push('-m', 'models/ggml-' + WHISPER_MODEL_SIZE + '.bin');
  }
  // Medical vocabulary hint via initial prompt
  args.push('--prompt', 'Medical patient encounter. Pediatric. Clinical terms, diagnoses, medications.');
  return args;
 }
 /**
 * Parse whisper output into clean text
 */
 function parseOutput(stdout) {
  if (!stdout) return '';
  // whisper.cpp with --no-timestamps outputs plain text
  // whisper.cpp with timestamps: [00:00:00.000 --> 00:00:05.000]   Hello world
  var lines = stdout.split('\n');
  var text = [];
  for (var i = 0; i < lines.length; i++) {
    var line = lines[i];
    // Strip timestamp prefix if present
    var match = line.match(/^\[[\d:.]+\s*-->\s*[\d:.]+\]\s*(.*)$/);
    if (match) {
      text.push(match[1].trim());
    } else if (line.trim() && !line.startsWith('whisper_') && !line.startsWith('main:')) {
      // Plain text line (skip whisper.cpp log lines)
      text.push(line.trim());
    }
  }
  return text.join(' ');
 }
 /**
 * Clean up temp files
 */
 function cleanup(tmpFile, wavPath) {
  try { fs.unlinkSync(tmpFile); } catch(e) {}
  if (wavPath) { try { fs.unlinkSync(wavPath); } catch(e) {} }
 }
 module.exports = { transcribeWithLocal, isLocalWhisperConfigured };
--- a/src/utils/ttsGoogle.js
+++ b/src/utils/ttsGoogle.js
@ -1,47 +0,0 @@
 // ============================================================
 // TTS GOOGLE — Google Cloud Text-to-Speech (direct, no LiteLLM)
 // Uses google-auth-library (transitive dep of @google-cloud/vertexai)
 // Voices: en-US-Journey-F (female), en-US-Journey-D (male),
 //         en-US-Studio-O, en-US-Neural2-C, etc.
 // Requires: GOOGLE_VERTEX_PROJECT, ADC or GOOGLE_APPLICATION_CREDENTIALS
 // ============================================================
 async function synthesizeWithGoogleTTS(text, voiceName) {
  var GoogleAuth = require('google-auth-library').GoogleAuth;
  var axios = require('axios');
  var auth = new GoogleAuth({
    scopes: ['https://www.googleapis.com/auth/cloud-platform']
  });
  var client    = await auth.getClient();
  var tokenData = await client.getAccessToken();
  var token     = tokenData.token;
  var voice     = voiceName || process.env.GOOGLE_TTS_VOICE || 'en-US-Journey-F';
  var langCode  = voice.substring(0, 5); // e.g. en-US
  var resp = await axios.post(
    'https://texttospeech.googleapis.com/v1/text:synthesize',
    {
      input: { text: text },
      voice: { languageCode: langCode, name: voice },
      audioConfig: { audioEncoding: 'MP3' }
    },
    {
      headers: {
        'Authorization': 'Bearer ' + token,
        'Content-Type': 'application/json'
      },
      timeout: 30000
    }
  );
  return Buffer.from(resp.data.audioContent, 'base64');
 }
 function isGoogleTTSConfigured() {
  if (!process.env.GOOGLE_VERTEX_PROJECT) return false;
  try { require('google-auth-library'); return true; } catch(e) { return false; }
 }
 module.exports = { synthesizeWithGoogleTTS, isGoogleTTSConfigured };
--- a/src/utils/ttsProvider.js
+++ b/src/utils/ttsProvider.js
@ -1,18 +1,16 @@
 const { isGoogleTTSConfigured } = require('./ttsGoogle');
 const { getLiteLLMHeaders } = require('./litellm');
-const GOOGLE_TTS_VOICES = ['en-US-Journey-F','en-US-Journey-D','en-US-Studio-O','en-US-Neural2-A','en-US-Neural2-C','en-US-Neural2-D','en-US-Neural2-E','en-US-Neural2-F','en-US-Neural2-G','en-US-Neural2-H','en-US-Neural2-I','en-US-Neural2-J'];
+function parseList(value) {
-const OPENAI_TTS_VOICES = ['alloy','echo','fable','onyx','nova','shimmer','ash','sage','coral'];
+  return String(value || '')
-const ELEVENLABS_DEFAULT_VOICES = ['pNInz6obpgDQGcFmaJgB'];
+    .split(',')
    .map(function(item) { return item.trim(); })
    .filter(Boolean);
 }
 function getTTSProvider() {
  var env = process.env.TTS_PROVIDER;
  if (env === 'google') return 'google';
  if (env === 'litellm') return 'litellm';
  if (env === 'elevenlabs') return 'elevenlabs';
  if (process.env.LITELLM_API_BASE) return 'litellm';
  if (isGoogleTTSConfigured()) return 'google';
  if (process.env.ELEVENLABS_API_KEY) return 'elevenlabs';
  return 'none';
 }
@ -22,9 +20,7 @@ function getTTSEnvProvider() {
 function getTTSVoiceLists() {
  return {
-    google: GOOGLE_TTS_VOICES,
+    litellm: parseList(process.env.LITELLM_TTS_VOICES)
    litellm: OPENAI_TTS_VOICES,
    elevenlabs: ELEVENLABS_DEFAULT_VOICES
  };
 }
@ -57,16 +53,13 @@ function getLiteLLMTTSDiscoveryItems(models, opts) {
  if (opts.currentVoice) {
    pushUniqueTTSItem(items, { id: opts.currentVoice, name: opts.currentVoice, source: 'configured-voice', kind: 'voice' });
  }
-  OPENAI_TTS_VOICES.forEach(function(voice) {
+  getTTSVoiceLists().litellm.forEach(function(voice) {
-    pushUniqueTTSItem(items, { id: voice, name: voice, source: 'openai-compatible', kind: 'voice' });
+    pushUniqueTTSItem(items, { id: voice, name: voice, source: 'configured-voice-list', kind: 'voice' });
  });
  return items;
 }
 module.exports = {
  ELEVENLABS_DEFAULT_VOICES,
  GOOGLE_TTS_VOICES,
  OPENAI_TTS_VOICES,
  getTTSEnvProvider,
  getLiteLLMTTSDiscoveryItems,
  getLiteLLMHeaders,
--- a/test/stt-provider.test.js
+++ b/test/stt-provider.test.js
@ -24,14 +24,12 @@ test('STT provider respects explicit provider override', () => {
  });
 });
-test('STT provider auto-detects in existing priority order', () => {
+test('STT provider only auto-detects LiteLLM', () => {
  const sttProvider = require('../src/utils/sttProvider');
  withEnv({}, () => {
-    assert.equal(sttProvider.getSTTProvider({ google: true, aws: true, local: true, litellm: true, openai: true }), 'google');
+    assert.equal(sttProvider.getSTTProvider({ google: true, aws: true, local: true, litellm: true, openai: true }), 'litellm');
    assert.equal(sttProvider.getSTTProvider({ aws: true, local: true, litellm: true, openai: true }), 'aws');
    assert.equal(sttProvider.getSTTProvider({ local: true, litellm: true, openai: true }), 'local');
    assert.equal(sttProvider.getSTTProvider({ litellm: true, openai: true }), 'litellm');
-    assert.equal(sttProvider.getSTTProvider({ openai: true }), 'openai');
+    assert.equal(sttProvider.getSTTProvider({ openai: true }), 'none');
    assert.equal(sttProvider.getSTTProvider({}), 'none');
  });
 });
@ -39,11 +37,7 @@ test('STT provider auto-detects in existing priority order', () => {
 test('STT model lists preserve admin defaults', () => {
  const sttProvider = require('../src/utils/sttProvider');
  assert.deepEqual(sttProvider.getSTTModelLists(), {
-    google: ['gemini-2.0-flash', 'gemini-2.5-flash', 'gemini-2.0-flash-lite'],
+    litellm: ['local-whisper-large-v3-turbo', 'local-whisper-large-v3', 'local-parakeet-v3', 'local-whisper-1', 'gemini-2.0-flash', 'gemini-2.5-flash']
    aws: ['standard', 'medical'],
    openai: ['whisper-1'],
    litellm: ['local-whisper-large-v3-turbo', 'local-whisper-large-v3', 'local-parakeet-v3', 'local-whisper-1', 'gemini-2.0-flash', 'gemini-2.5-flash'],
    local: ['tiny', 'base', 'small', 'medium', 'large']
  });
 });
--- a/test/tts-provider.test.js
+++ b/test/tts-provider.test.js
@ -1,7 +1,7 @@
 const { test } = require('node:test');
 const assert = require('node:assert/strict');
-const ENV_KEYS = ['TTS_PROVIDER', 'LITELLM_API_BASE', 'LITELLM_API_KEY', 'ELEVENLABS_API_KEY', 'GOOGLE_VERTEX_PROJECT'];
+const ENV_KEYS = ['TTS_PROVIDER', 'LITELLM_API_BASE', 'LITELLM_API_KEY', 'LITELLM_TTS_VOICES'];
 function withEnv(overrides, fn) {
  const previous = {};
@ -25,27 +25,22 @@ test('TTS provider defaults to LiteLLM when gateway is configured', () => {
  });
 });
-test('TTS provider respects explicit provider override', () => {
+test('TTS provider ignores non-LiteLLM provider overrides', () => {
  const ttsProvider = require('../src/utils/ttsProvider');
  withEnv({ TTS_PROVIDER: 'elevenlabs', LITELLM_API_BASE: 'https://llm.example.com' }, () => {
    assert.equal(ttsProvider.getTTSEnvProvider(), 'elevenlabs');
-    assert.equal(ttsProvider.getTTSProvider(), 'elevenlabs');
+    assert.equal(ttsProvider.getTTSProvider(), 'litellm');
  });
 });
-test('LiteLLM TTS voice list stays OpenAI-compatible', () => {
+test('LiteLLM TTS voice list comes from configured voice catalog', () => {
  const ttsProvider = require('../src/utils/ttsProvider');
-  assert.deepEqual(ttsProvider.getTTSVoiceLists().litellm, [
+  withEnv({ LITELLM_TTS_VOICES: 'sherpa/kokoro:am_adam, sherpa/kokoro:af_bella' }, () => {
-    'alloy',
+    assert.deepEqual(ttsProvider.getTTSVoiceLists().litellm, [
-    'echo',
+      'sherpa/kokoro:am_adam',
-    'fable',
+      'sherpa/kokoro:af_bella'
-    'onyx',
+    ]);
-    'nova',
+  });
    'shimmer',
    'ash',
    'sage',
    'coral'
  ]);
 });
 test('TTS model discovery uses audio_speech metadata only', () => {
@ -79,20 +74,22 @@ test('LiteLLM TTS model extraction filters gateway model objects', () => {
 test('LiteLLM TTS discovery includes metadata models and configured fallbacks', () => {
  const ttsProvider = require('../src/utils/ttsProvider');
-  const items = ttsProvider.getLiteLLMTTSDiscoveryItems([
+  withEnv({ LITELLM_TTS_VOICES: 'sherpa/kokoro:am_adam,sherpa/kokoro:af_bella' }, () => {
-    { model_name: 'local-kokoro-tts', model_info: { mode: 'audio_speech' } },
+    const items = ttsProvider.getLiteLLMTTSDiscoveryItems([
-    { model_name: 'not-tts-by-name-only' },
+      { model_name: 'local-kokoro-tts', model_info: { mode: 'audio_speech' } },
-    { model_name: 'local-parakeet-v3', model_info: { mode: 'audio_transcription' } }
+      { model_name: 'not-tts-by-name-only' },
-  ], {
+      { model_name: 'local-parakeet-v3', model_info: { mode: 'audio_transcription' } }
-    currentModel: 'local-kokoro-tts',
+    ], {
-    currentVoice: 'sherpa/kokoro:am_adam'
+      currentModel: 'local-kokoro-tts',
      currentVoice: 'sherpa/kokoro:am_adam'
    });
    assert.deepEqual(items.slice(0, 3), [
      { id: 'local-kokoro-tts', name: 'local-kokoro-tts', source: 'gateway-api', kind: 'model' },
      { id: 'sherpa/kokoro:am_adam', name: 'sherpa/kokoro:am_adam', source: 'configured-voice', kind: 'voice' },
      { id: 'sherpa/kokoro:af_bella', name: 'sherpa/kokoro:af_bella', source: 'configured-voice-list', kind: 'voice' }
    ]);
    assert.equal(items.some(function(item) { return item.id === 'not-tts-by-name-only'; }), false);
  });
  assert.deepEqual(items.slice(0, 3), [
    { id: 'local-kokoro-tts', name: 'local-kokoro-tts', source: 'gateway-api', kind: 'model' },
    { id: 'sherpa/kokoro:am_adam', name: 'sherpa/kokoro:am_adam', source: 'configured-voice', kind: 'voice' },
    { id: 'alloy', name: 'alloy', source: 'openai-compatible', kind: 'voice' }
  ]);
  assert.equal(items.some(function(item) { return item.id === 'not-tts-by-name-only'; }), false);
 });
 test('LiteLLM TTS discovery still shows configured model if metadata lookup fails', () => {