Your earbuds are brilliant at one thing: playing music into your ears. They were never really designed for the other direction – getting your voice out.

On a phone call, that's mostly fine. The person on the other end fills in the gaps. They hear "...ing at six" and know you said "meeting at six." A speech model is far less forgiving. It needs every syllable, and through earbuds it gets a surprisingly bad copy of them.

We learned this the hard way. We were early, heavy users of dictation apps, and they kept getting our words wrong whenever we spoke through earbuds. Friends with AirPods had the same problem. Then we noticed the dictation apps themselves tell you not to use AirPods. Wispr Flow flags them in its mic picker with a warning that they may be slow to start and make more errors, and recommends your laptop's built-in mic instead.

So what goes wrong between your mouth and the model? Three things.

The pipe gets narrow. Classic Bluetooth has two modes. One streams rich stereo to your ears but carries no microphone. The moment any app opens the mic, the connection drops to the hands-free profile: mono audio at 8 or 16 kHz, in both directions. That's why your music suddenly sounds like an old phone call when you start dictating. At 8 kHz, nothing above 4 kHz survives – and that's where consonants like "s," "f" and "th" live. Doubling the sample rate from 8 to 16 kHz alone has been shown to cut word error rate (WER) by about 20%, which is why dictation systems are built for 16 kHz or more.

The mic is in the wrong place. An earbud mic sits by your ear, facing the world, because its main job is hearing noise so ANC can cancel it. Your mouth is an afterthought. Distance matters a lot: OpenAI's Whisper makes more than twice as many errors on meeting audio from a single distant mic (36.4% WER) as on the same meetings through headset mics (16.9%).

The cleanup does damage. To make up for both, earbuds run aggressive noise suppression, tuned to sound fine to a human on a call. But suppression leaves smeared, muffled bits that are neither your voice nor the noise – and research shows those artifacts, not the leftover noise, are the main reason speech recognition gets worse. Speak softly or sit in a busy café and the algorithm starts eating your words along with the background.

We put it together and, in our own tests, AirPods hit a WER of around 34%. That's one word in three wrong. And this isn't an AirPods problem. It's what happens to any pair of earbuds tuned for music.

Why this matters for voice AI?

Speech models are good enough now. On clean audio, talking is as accurate as typing and much faster. The bottleneck has moved from the model to the hardware in your ears.

And we're about to talk to computers a lot more – dictating to agents, briefing them, correcting them mid-task, in cafés, open offices, cars and trains. If the most used voice AI apps don't work on the most used voice hardware, the problem isn't the model. It's the interface.

That's why we built project mnemosyne's earphones for speaking first. A voice pickup sensor reads your voice through bone vibration, and directional mics pointed at your mouth do the rest – so the model hears you, not the room, without the muffled sound that usually comes with isolation.

Built for music, earbuds are great. Built for your voice, they can be so much more.