"In a 911 call, every misheard word could be a missed address, a wrong turn, a life lost."
Trace911 is a real-time AI dispatcher assistant that cleans noisy emergency call audio, transcribes it live, analyzes urgency and sentiment, and automatically dispatches help — so dispatchers can always hear what matters, even when everything is chaos.
Built in 48 hours at a hackathon using ai-coustics, Gradium, and Google Gemini.
Voice agents and dispatch systems are built and tested in near-perfect acoustic conditions. But real 911 calls are chaotic — background noise, panicked voices, bad connections, multiple people talking at once. Dispatchers mishear critical information. Addresses get lost. Response times suffer.
Trace911 sits between the caller and the dispatcher. It:
- Cleans the incoming audio in real time using ai-coustics noise cancellation
- Transcribes the caller's voice live using Gradium STT
- Analyzes urgency, sentiment, location and emergency type using Google Gemini
- Speaks critical alerts to the dispatcher via Gradium TTS
- Auto-dispatches help when enough information is confirmed
We designed a Word Error Rate (WER) comparison to prove ai-coustics works:
| Word Error Rate | |
|---|---|
| Without ai-coustics | 34.2% |
| With ai-coustics | 3.1% |
| Accuracy gain | 91% |
Tested on a real 911 fire emergency call recording.
The dispatcher opens Trace911 and presses Demo Call:
- Enhanced audio streams live
- Transcript appears word by word in real time
- AI analyzes every 15 seconds — urgency, sentiment, location, emergency type
- When urgency hits CRITICAL, TTS speaks: "Critical. Fire. Route 7 Abington."
- When location is confirmed, system auto-dispatches: "Units dispatched to Route 7 Abington."
- Dispatcher presses Reveal Metrics to show the WER improvement
| Component | Technology |
|---|---|
| Audio noise cancellation | ai-coustics Quail Voice Focus |
| Live transcription | Gradium STT (WebSocket, 24kHz PCM) |
| Real-time analysis | Google Gemini 2.5 Flash |
| Dispatcher voice alerts | Gradium TTS |
| API server | Python Flask |
| Dashboard | React + Tailwind (Lovable) |
| WER measurement | editdistance Python library |
Add these to calls/.env:
GRADIUM_API_KEY=your_key
GEMINI_API_KEY=your_key
AIC_SDK_LICENSE=your_keycd calls
pip install -r requirements.txt
python server.pyAPI runs on http://localhost:5000
cd frontend
npm install
npm run devDashboard opens at http://localhost:8080
- Open http://localhost:8080
- Press Play Original — hear the noisy 911 call
- Press Demo Call — watch Trace911 work in real time
- Press Reveal Metrics — see the WER improvement
Trace911/
├── calls/ Working demo pipeline
│ ├── server.py Flask API (6 endpoints)
│ ├── stream_transcribe.py Gradium STT streaming
│ ├── analyze.py Gemini real-time analysis + TTS alerts
│ ├── speak.py Gradium TTS module
│ ├── wer.py Word Error Rate measurement
│ └── clean_audio.py ai-coustics noise cancellation
├── frontend/ React dashboard (Lovable)
├── backend/ Production FastAPI architecture (designed, not yet wired)
└── docs/ Production API contracts, WebSocket events, data modelThe calls/ folder is the working hackathon demo. The backend/ and docs/ folders contain the production-ready architecture designed for real deployment via Telnyx PSTN infrastructure.
In production Trace911 would:
- Receive real 911 calls via Telnyx PSTN infrastructure
- Route audio through LiveKit for real-time streaming
- Use the FastAPI backend (see backend/) with SQLite for call history
- Support multiple simultaneous dispatch stations
- Integrate with CAD (Computer Aided Dispatch) systems
- ai-coustics — Audio enhancement SDK
- Gradium — Voice AI STT + TTS
- Google Gemini — Multimodal AI analysis
- Lovable — Frontend generation
- Flask — API server
Built at a hackathon in 48 hours.