VoiceTriage Review
A Next.js app that uses Whisper STT, Llama 3.3, and browser TTS to conduct a voice-first symptom triage with red-flag detection.
Each review covers deployability, value versus commercial alternatives, and privacy model. Tools that can run locally were started and exercised; mobile or backend-dependent tools were assessed from published builds, source code, and deploy guides. Ratings reflect what we were able to verify.
VoiceTriage — Voice-First Clinical Intake Agent
A Next.js app that uses Whisper STT, Llama 3.3, and browser TTS to conduct a voice-first symptom triage with red-flag detection. Last updated: 2026-06-21.
One-sentence verdict: A research demo with strong safety scaffolding and clear documentation, but explicitly not a substitute for professional triage or emergency services.
What the System Is
VoiceTriage is a Next.js 15 application that collects symptoms through voice (or text), asks focused follow-up questions, and produces a structured intake summary. The pipeline uses Groq's Whisper for transcription, Groq's Llama 3.3 70B for reasoning, and the browser SpeechSynthesis API for text-to-speech. A deterministic regex scanner runs in parallel with the LLM to catch red-flag patterns the model might miss.
The deploy guide reports the service was started on port 5007.
| Key data | |
|---|---|
| Category | Symptom Checker / Clinical Intake |
| Language | TypeScript / Next.js |
| License | Apache 2.0 |
| Self-hosted | Yes |
| AI provider | Groq (Whisper + Llama 3.3 70B) |
| Database | None; conversation state is client-side |
| Auth | None |
| Tests | Vitest unit tests for red-flag detection; persona eval harness |
How to Install and Deploy
cd /data2/docker/going-global/repos/voice-triage
cp .env.example .env.local
# Add GROQ_API_KEY
npm install
PORT=5007 npm run dev
Then open http://localhost:5007 in a Chromium-based browser and allow microphone access.
How to Test
The documented test flow is:
- Open
http://localhost:5007. - Click the microphone button and describe a symptom.
- Verify the agent transcribes the input, asks a follow-up, and speaks the reply.
- Continue up to 6 turns or until a red flag triggers an escalation banner.
- Click to generate the structured intake summary and download it as JSON.
- Run
npm testfor the deterministic safety scanner unit tests. - Run
npm run evalto replay 8 synthetic patient personas against the live API.
Privacy & Compliance
VoiceTriage is not HIPAA compliant. It stores no data on a server, but every voice utterance is sent to Groq for transcription and the LLM for reasoning. The README explicitly labels the project as a "research demo only" and warns against using it for real triage. Do not enter health information you are not comfortable sending to Groq.
VoiceTriage vs Commercial Symptom Checkers
| Dimension | VoiceTriage | Commercial Checkers (e.g., Ada Health, K Health, WebMD) |
|---|---|---|
| Cost | Free / self-hosted + Groq usage | Freemium or subscription |
| Medical advice | Explicitly not for real triage | Clinically validated algorithms |
| Red-flag detection | LLM + deterministic regex scan | Clinically curated rule sets |
| Data storage | No server-side persistence | Cloud history tied to account |
| Voice interface | Browser-native mic + TTS | Varies; usually text-first |
| Evidence base | None claimed | Often published validation studies |
| EMR integration | None | Some offer provider connectivity |
| Customization | Full source code | Closed |
Who Should Use It
- Developers building voice-first triage or intake prototypes.
- Researchers studying LLM safety patterns and red-flag detection.
- Teams that want a well-documented example of Whisper + Llama + browser TTS integration.
Who Shouldn't Use It
- Anyone seeking medical triage or emergency guidance; the README tells users to call emergency services instead.
- Organizations needing a clinically validated, regulated symptom checker.
- Users who want persistent health records or multi-session history.
FAQ
Is VoiceTriage a medical device?
No. The README states explicitly that it is a research demo and must not be used for real triage.
Does it store my conversation?
No server-side database is used. Conversation state lives in the browser's React state and is lost when the tab is closed.
What happens if the LLM misses a red flag?
A deterministic regex scanner in lib/safety.ts runs on every user transcript in parallel with the LLM. The UI escalates if either signal fires.
Verdict
VoiceTriage is a compact, well-architected research demo. Its two-layer safety design (LLM + deterministic regex), clear README, and eval harness make it a better learning scaffold than most prototypes. It is not, and does not claim to be, a replacement for clinically validated triage tools.
Ratings: Deployability 4/5 · Value vs Commercial 2/5 · Privacy Compliance 2/5