Subtle: earbuds trained to hear a whisper over a crowd
A small Stanford-founded startup builds earbuds whose AI model isolates the wearer's own voice instead of blocking noise, catching speech below a whisper or across a loud room.
I came across this while reading old CES coverage and one detail stuck with me: a $199 pair of earbuds picked up a whisper so quiet that the journalist testing them over a video call couldn't hear it through the audio feed, yet the earbuds transcribed it word for word.
The company is Subtle, a small voice AI startup founded by a group of Stanford engineers who spent most of 2025 in stealth, on a modest six million dollar seed round, before showing their first piece of hardware, the Voicebuds, at CES in January. Most earbuds treat noise as something to erase: active noise cancellation listens to everything around you and subtracts it, which is fine for music but falls apart the moment you actually need to be heard, either because the room is too loud or because you don't want to be overheard at all.
That's the clever part. Subtle flipped the problem: rather than modeling the noise, its AI model was trained to recognize the wearer's own voice specifically and isolate that signal from everything else, using a multi-microphone array feeding a custom on-device chip. The result is an earbud tuned for the two extremes a normal microphone struggles with: a whisper quieter than a library voice, or a CES show floor loud enough to drown out a conversation. Subtle claims five times fewer transcription errors than Apple's AirPods Pro 3 running OpenAI's own transcription, measured in exactly those conditions.
Eight months later, this isn't just a demo. The first shipment sold out, and the second wave, now priced at 249 dollars, ships this fall. The earbuds are still Apple-only and lean on a subscription for the more advanced dictation features, so they won't replace anyone's AirPods overnight. But they're a good reminder that in a market obsessed with noise cancellation and battery life, there was still an obvious axis nobody had optimized for: making sure the microphone actually hears you.