A mic array and a speech enhancement model is not the answer to robotics audio.
That's the default today: bolt a microphone array onto the robot, run a model that cleans up voices, and call it hearing. It was built for smart speakers and conference calls, not for machines that move through the world. Here is what we believe robots need instead.
- 01
Hearing is more than speech
Speech enhancement treats everything that isn't a voice as noise to throw away. For a robot, that noise is often the signal: an alarm, a grinding bearing, footsteps behind it, glass breaking in the next aisle.
Sound carries around corners, through walls, and in the dark, where cameras and lidar see nothing. A robot needs to hear all of it, and know which sounds matter.
- 02
Robots need answers, not cleaner audio
An enhancement model outputs a cleaner waveform, built for human ears or a transcription engine. A robot can't act on a waveform.
It needs to know what happened, where it came from, and when, as structured events it can plan around.
- 03
Senses only add up when they share a clock
Audio alone is one sense. Video alone is another. Each runs on its own timeline and tells only part of the story. A robot that can't line them up can't tell whether the clang it just heard came from its own arm or from something across the room.
A truly perceptive system correlates its senses: every sensor measures the same moment, and their data is shared and connected. That is why we are adding a shared clock: hardware-level time sync between Overtone and your robot, so every sound lines up exactly with camera, lidar, IMU, and servo data.
- 04
Hearing has to work in motion
Mic arrays and enhancement models are tuned for a device sitting still on a table in a quiet room. Robots walk, drive, turn, and vibrate. Their motors hum, their joints whine, and the array moves with the body.
We design for the noisy, moving, real world first, and treat a quiet lab as the easy case.
- 05
Hearing shouldn't cost the robot anything
Every cycle on a robot's CPU and GPU is already spoken for by navigation, vision, and planning. Running audio models there means something else gets worse.
So Overtone brings its own compute. Your robot gets clean, structured events about what it heard and where, with zero added load on the processors it already depends on.
- 06
What a robot hears should stay with the robot
Robots work in homes, hospitals, and factories, where people talk freely. Streaming raw audio to the cloud to make sense of it is a privacy problem waiting to happen.
Overtone processes sound on the module. Raw recordings never have to leave it. Only the meaning does.
- 07
Precision is earned, not approximated
We come from professional audio: building hardware, producing music, and designing surround sound and room acoustics. In that world, a few milliseconds or a few degrees is the difference between right and wrong.
We hold machine hearing to the same standard. Timing, direction, and clarity are the product, not a spec sheet footnote.
- 08
Built with the people who use it
Nobody knows what a robot needs to hear better than the teams deploying them. That's why we're building Overtone alongside a small group of design partners, not in isolation.
Their robots, their environments, and their edge cases shape every decision we make.
Teach your robots to hear
If you believe this too, we'd like to build it with you.