Hackathon · Voice AI Hack · London 2026
TrueVoice
Patients minimise. Voices don't.
A platform that listens during medical consultations and flags when a patient sounds worse than their words say.
Overall Winner - Voice AI Hack London 2026
First place across the full event in the Voice & Medical track sponsored by Thymia and Speechmatics.
- Event
- Voice AI Hack, London
- Result
- Overall Winner
- Stack
- Python, FastAPI, WebRTC, Next.js, Claude
- Team
- Joan Torres Gordo and Indigo Luksch
Why it matters
People often tell their doctor they are fine when they are not. The words go in the notes, but the strain in their voice is easy to miss in a short appointment. TrueVoice listens to both while the consultation happens, and points out the moments where they disagree.
How it works
Patient and doctor join a video call, each from their own device.
Video call
Each device captures once, uses it twice
WebRTC · STUN and TURN
Connecting…
Laptop mic
One mic on the desk for both voices
audio only
Backend
Sends every frame two ways
FastAPI · 16 kHz · 40 ms frames
Speechmatics
Live medical transcript
who spoke: by device, or by voice
"I'm fine, honestly."
Thymia
Signals in the voice
Helios · Apollo · Psyche
distress: high
Mismatch check
Both within a minute
- phrase matched
- voice over threshold
Doctor's screen
A live flag with a note
Claude Haiku · 2 s limit
Says fine, sounds distressed
Report
When the doctor asks
Claude Sonnet
Summary ready
Decisions we made
Our own call, not a video app
For telehealth, each browser captures its microphone once and uses it twice: one copy goes into a peer-to-peer video call, the other goes straight to the voice analysis as clean 16 kHz audio from the speaker's own device.
Trade-off. We had to build the call ourselves, including the signalling and a relay server so it connects through home routers. It cost us a lot of time, and very nearly the submission.
Rules raise the flag, the model only words it
A flag fires only when a phrase like "I'm fine" turns up and a voice signal crosses its threshold within the same minute. Claude only writes the one-line note, with a two-second limit and a plain sentence to fall back on.
Trade-off. Easy to explain and test, but a fixed list of 23 English phrases misses anything said another way.
A fast model live, a stronger one after
Claude Haiku writes the notes during the appointment, where a slow note is a useless one. Claude Sonnet writes the summary only when the doctor asks for it.
Trade-off. The live notes stay short and plain, and the depth goes into the report.
The last ten minutes
Minutes before the deadline, the video call stopped working. We spent the rest of the time fixing it and got it connecting again with about five minutes left. We still had not recorded the demo, so we recorded it from the script we had ready, uploaded it, and submitted with about a minute to spare. We won overall, and the call we nearly lost was a big part of why.
Photos
Gallery
The full engineering is on GitHub.