Mission WHISPER
AI Context Device
Context is limited by what a model can see today. Meeting assistants like meetjamie.ai or Granola only work when you actively use them. Other assistant devices have failed in the past — friend.com, Humane — because they tried to do too much. What matters, in a simpler form, is just getting context from in-person conversations. Constantly.
Your mission: build a simple wearable device with a mic that stores voice embeddings from the user as a voice fingerprint. It checks locally whether what the device is hearing through the mic is from the user themselves (not someone else's voice). If it is, transcribe that audio — on device, ideally — and push it to a personal place of the user's choosing (a GitHub repo, Google Drive, whatever).
Think a wrist band like Whoop. A small companion app on the phone to pair the device. Nothing fancy — just solid engineering on a hard problem.
Objectives
- Use voice embeddings (e.g. Qwen3-Voice-Embedding-12Hz-0.6B on HuggingFace) to build a local voice fingerprint. The device should know when it's hearing the wearer vs. someone else.
- Transcribe speech on-device (or as close to on-device as you can get). Accuracy matters more than speed.
- Push transcripts somewhere the user controls. GitHub repo, Google Drive, local folder, whatever they prefer.
- Build a companion phone app to pair the device and review what it captured.
- Make it something you'd actually wear. Wrist band, clip, pendant. Not a science fair project.
Constraints
- Target production cost at scale: ~€50 per unit
- Battery should last a full day of normal use. Nobody wants to charge a wristband at lunch.
- Transcription has to be good, even in noisy rooms. If it can't handle a coffee shop, it's useless.
- It should look like something a normal person would wear. Not a dev kit with exposed wires.
- All voice processing stays local. No audio leaves the device unencrypted or without the user saying so.
Resources
CLASSIFIED BRIEFING
Scroll to activate countdown.