Health screening from 30 seconds of voice.
Mental wellbeing, respiratory health, and cognitive performance — no hardware, no questionnaires, no downloads.
The WHO estimates that 75% of depression cases in middle-income countries go undetected. The gold-standard screening tool (PHQ-9) assumes the patient already knows they're unwell, has the words to describe it, and is willing to write it down. For most people, that's the wall that keeps them out of the consultation in the first place.
Echo operates before that wall. Thirty seconds of voice. Twelve acoustic biomarkers. Three actionable scores. When the indicators climb, the report itself recommends a referral.
- You speak for 30 seconds about anything — your day, your week, what's on your mind.
- Echo extracts 12 acoustic features in real time with Meyda, 100% in the browser.
- The transcript appears live via the Web Speech API; when system audio is shared (video call, podcast), Whisper takes over.
- GPT-5-mini with a strict JSON Schema interprets the features + transcript against published clinical thresholds.
- Three scores with visible justification — plus a Clinical mode that exposes all 12 features with their normality bands and source references.
echo-screening.vercel.app — Chrome or Edge on desktop.
A close friend of mine has social phobia. I go with him to every hospital visit. That's where the real problem becomes obvious: the people who need help the most rarely reach the first appointment. Not because they don't want it. Because the first contact is expensive, intimidating, slow — and it asks the patient to recognize something is wrong before they have anything objective to point to.
Echo isn't a replacement for the doctor. It's the nudge that gets someone through the door before they can articulate what they're feeling.
The biomarkers Echo measures aren't speculative. They come from three lines of public, citable research:
| Dimension | Markers | Source |
|---|---|---|
| Mental wellbeing | jitter, shimmer, flat prosody, pause ratio | Quatieri et al., MIT Lincoln Lab — INTERSPEECH 2012. Depression alters laryngeal motor control through basal ganglia changes — before any visible symptoms appear. |
| Respiratory health | shimmer, spectral clarity, phonation | Sonde Health — validation on 1.2M voice samples, Frontiers in Psychiatry, 2024. |
| Cognitive performance | speech rate, lexical richness, pause ratio | MIT CSAIL with the Framingham cohort. Verbal fluency declines before the memory symptoms of mild cognitive impairment become visible. |
Echo turns that research into product. It doesn't invent it.
[Browser]
├─ getUserMedia ─────────── microphone
└─ getDisplayMedia (opt) ── system audio (video call, podcast)
│
▼
AudioContext + Meyda 5.x
│
┌────────┼────────┐
▼ ▼ ▼
live freq 12 features MediaRecorder
indicators client-side │
│ ▼
│ [POST /api/transcribe]
│ │
│ Whisper-1
│ │
└────────────┴───────► [POST /api/analyze]
│
GPT-5-mini
structured outputs
(strict JSON Schema)
│
▼
[Frontend renders]
├─ 3 scores with indicators
├─ Cross insights
├─ Clinical mode (12 features + bands)
├─ Temporal evolution
├─ Side-by-side comparison
└─ Audio replay
- Frontend: Next.js 16 (App Router), TypeScript, Tailwind v4, shadcn/ui, Recharts
- In-browser DSP: Meyda 5.x — RMS, ZCR, spectral centroid/flatness/rolloff, perceptual sharpness/spread, loudness, energy
- Speech-to-text: Web Speech API for the live transcript, Whisper for system audio
- AI: OpenAI GPT-5-mini with structured outputs (strict JSON Schema)
- Deploy: Vercel
- Raw audio is never persisted on the backend.
- Acoustic features are computed 100% on the client with Meyda.
- What travels to the server: the blob for one-off Whisper transcription, the numeric features, and the text transcript.
- No database. No persistent session. Each analysis is ephemeral.
- For regulated production (HIPAA / GDPR / Habeas Data Colombia), end-to-end encryption and a BAA with the AI provider are required — that's on the roadmap, not in this demo.
Worth knowing before you judge the demo:
- The evolution chart (5 days) shows simulated data. I left it visible on purpose, to communicate the recurring-use vision: Echo is longitudinal, not one-shot. The clinically meaningful signal is the delta between baseline and follow-up — not any single value.
- No formal clinical validation yet. The features and thresholds come from the literature. Validation against real cohorts is the next phase, not this one.
- Echo is screening, not diagnosis. When indicators escalate to
moderateorelevated, the report itself recommends referral to a professional. It never replaces a consultation. - Built in roughly 10 hours during a hackathon. The pipeline is real and auditable in this repository.
| Channel | Who pays | Why |
|---|---|---|
| B2C | End user (freemium) | In-browser self-screening; refers out to a professional when the signals warrant it. |
| B2B clinical | Clinics, EPS, insurers | The doctor activates Echo during the consultation as a complementary objective biomarker. |
| B2B corporate | HR at large companies | Wellbeing program — early detection of burnout and cognitive load in employees. |
Initial target market: Latin America. An EPS spends orders of magnitude more on hospitalizations for undiagnosed depression than on a scaled vocal screening would ever cost.
- Product: validated condition-specific markers — vocal fatigue, early Parkinson's, mild cognitive impairment.
- Model: fine-tuning over DAIC-WOZ (the academic standard dataset for depression) plus proprietary data from clinical pilots.
- Distribution: integration with WhatsApp Business, EPS portals, and corporate wellness apps. Zero downloads, zero friction.
- Compliance: registration as Software as a Medical Device class A (screening, not diagnosis) — INVIMA Colombia, FDA 510(k) United States.
git clone https://github.com/isjunrod/echo.git
cd echo
npm install
# OPENAI_API_KEY is the only required variable.
echo "OPENAI_API_KEY=sk-..." > .env.local
npm run dev
# → http://localhost:3000Tested in Chrome 124+ and Edge 124+ on macOS, Linux, and Windows. Safari requires the fallback (Web Speech API is limited).
MIT. Use the code. The "Echo" brand and the clinical proposition belong to the author.
Built by Junior Rodriguez at the Health & Human Optimization Hackathon — April 2026.