Inspiration
Talking is the fastest way to tell an AI what you want. But nobody wants to talk to their laptop out loud in a library, a meeting or a quiet train. So people type, or wait until they're alone. We wanted a way to talk to your computer in a whisper.
What it does
Murmur is a tiny microphone that sits in your mouth, small enough for a retainer. You whisper, and your words show up on your Mac as you say them. It listens only to your whisper. People talking right next to you are ignored. Murmur can type it into the app you have open, or pass it to an AI agent to act on.
How we built it
The microphone is on a small Vexo board and connected to the PC. Other people's voices are dropped before anything else hears them. Then it turns the whisper into text with Deepgram, running through Pipecat. ElevenLabs and Gemini iare the speech engines, and its voice can read replies back to you. Last, it shows your words on a simple page in the browser and passes each finished phrase on as a command.
Challenges we ran into
The hard part was hearing only the wearer. Our first version picked up everyone in the room. We tried loudness first. The mic is in your mouth, so you should be the loudest thing it hears. In a busy room that turned out to be false. Next we tried the board's motion sensor, hoping it would feel you speak, but a board in your mouth moves all the time. On a real recording, 99% of the whisper got through and none of the voices around it did. The other problem was lag. Murmur used to wait almost two seconds after you stopped before doing anything. We cut that to under one.
Accomplishments that we're proud of
It works on real hardware. You put the mic in your mouth, whisper, and read your words on the screen while the person beside you keeps talking. There is nothing to train on your voice and no wake word to remember, so anyone can put it in and use it.
What we learned
A whisper and a normal voice can be equally loud. What separates them is pitch. Our first two ideas looked right on paper and failed on the real mic. Forty seconds of real audio taught us more than our guesses did. And the less you ask of the user, the better. Dropping the wake word made it feel like talking.
What's next for Murmur
Make it fully wireless inside a real retainer. Then connect it to an agent that can use your computer, so whispering "text Sam I'm running late" gets it done. After that we want to test it in loud places like concerts and crowded streets.
Built With
- aiohttp
- bluetooth
- css
- deepgram
- elevenlabs
- gemini
- html
- javascript
- numpy
- pipecat
- python
- seeed-xiao-nrf52840
- vexo
- websockets
