Noise cancelling and voice isolation sound like the same thing. Both fight noise, and both use microphones. But they point in opposite directions.

ANC is about what goes into your ears. Voice isolation is about what comes out of your mouth. Most earbuds are excellent at the first and quietly mediocre at the second. For voice AI, the second is the one that matters.

How ANC works?

ANC listens to the world and plays back its opposite. A mic on the outside of the bud hears noise coming in. A mic on the inside hears whatever is left. A chip calculates the inverted wave, and the speaker plays it so the two cancel out before they reach your eardrum.

It's brilliant at long, steady, low sounds: plane engines, air conditioning, road rumble. It struggles with almost everything else. The chip has to respond before the sound wave moves on, which in practice limits ANC to frequencies below roughly 1–2 kHz. Higher, sharper sounds, including other people's voices, mostly slip through, and the ear tip's seal does the rest of the work.

And all of this happens on the way in. ANC changes what you hear. It does nothing for how you sound.

How voice isolation works?

Voice isolation solves the reverse problem. Out of everything the mics hear, keep your voice, drop the rest, and only then send it.

That's much harder. An earbud's mic sits a lot further from your mouth than a phone's does, so in a loud room the noise can easily be louder than you. And in a café, the noise is other people talking – which sounds exactly like you.

The best earbuds attack this with three tools.

Beamforming. Several mics combine their signals to favor sound arriving from the direction of your mouth and tune down sound from everywhere else.

A voice pickup unit. A bone-conduction sensor that feels the vibration of your head when you speak and ignores sound traveling through the air. It tells the earbud exactly when you're talking and what your voice looks like, even when the room is louder than you.

Neural noise reduction. A model fuses all of those signals and rebuilds your voice without the background.

It works, but it's a balancing act. Bone conduction captures a duller version of your voice than air does, and a whisper barely vibrates your skull at all. Lean on the sensor too hard and you come out sounding muffled, or not at all.

Why this matters for voice AI?

The box says "noise cancelling," so people assume it covers both directions. It doesn't. Even earbud makers are starting to say so: one direction is about hearing the other person, the other is about them hearing you, and they're built with different hardware.

There's a second gap that almost nobody talks about. Even the best voice isolation today is tuned for a person on a phone call. It's scored on how good your speech sounds to human listeners. A speech model doesn't care how pleasant you sound. It cares whether every consonant made it through. Processing that sounds clean to a person can still smear exactly the details a model needs – which is why transcription falls apart on earbuds even when calls sound fine.

That's the gap we're building into. Project mnemosyne's earphones have ANC, because you should hear your agent clearly. But most of our work went the other way: directional mics aimed at your mouth, a voice pickup sensor, and processing tuned for the model on the other end, not just the ear.

ANC makes the world quieter for you. Voice isolation makes you clearer to the world. If you're going to talk to AI all day, you need both – and only one of them is on most spec sheets.