Earshot

Inspiration

We kept coming back to one quiet danger. A smoke alarm blares in an empty kitchen while someone sleeps just down the hall, unable to hear it. For millions of Deaf and hard of hearing people, sound-based warnings simply do not reach them, and the fixes cost a fortune. We wanted to build something that turns the sounds that matter into things you can see and feel, so safety is not a privilege.

What it does

Earshot listens for the sounds that matter in your home, like a smoke alarm, a doorbell, or a crying baby. The moment it hears one, it flashes a light and buzzes your wristband in about a second, so you always know what is happening around you. It even learns new sounds: record your doorbell or kettle three times and it remembers them. Everything runs on the device, so your audio never leaves the room.

How we built it

Earshot runs on a Raspberry Pi with a USB mic. Google's YAMNet model listens to live audio and flags sounds like alarms, doorbells, and knocks fully offline. A FastAPI backend turns each one into an LED flash, a buzz, a phone push, and a dashboard row within a second over WebSocket. We even trained our own alarm detector on YAMNet embeddings for higher accuracy and fewer false alarms.

Challenges we ran into

Our biggest challenges were physical and acoustic. Some hardware we planned on just was not available, so we reworked the wearable around the parts we could actually get our hands on. Then the model barely reacted to real sounds at first. Our mic gain was too low and our thresholds were off, so quiet or noisy audio slipped right past it. We fixed it by tuning the gain and thresholds against real room noise.

Accomplishments that we're proud of

We are proud that our detector actually works in a loud room, not just on clean clips. We trained our own alarm classifier on YAMNet embeddings, tuned it against real noise, and got it running fully offline on a Raspberry Pi. Seeing that tiny board hear a smoke alarm and fire off a light and a buzz in about a second, with no cloud and no internet, made the whole idea feel real.

What we learned

We learned that a model is only as good as the audio you feed it. Most of our real gains came from mic gain, window timing, and thresholds, not the network itself. We also learned to respect the Raspberry Pi's limits and keep every heavy step off the device, so training happened on a laptop and only a tiny artifact shipped to the Pi for fast offline detection.

What's next for Earshot

Next we want Earshot to learn more sounds and hear them more reliably. We would expand teach mode so users can add their own alerts in seconds, and train detectors for more critical sounds like sirens and alarms. We also want a real wearable microphone instead of one in a laptop, longer battery life, and directional hints so a user knows not just what made the sound, but where it came from.

Built With

• fastapi
• html
• javascript
• mongodb
• python
• raspberry-pi
• tensorflow
• typescript