Silent Speech

Watch on X →
16% WER on 1 hour of training data · this video has sound all the way through!

The idea is that it's simple to use: you keep your mouth closed and your lips still, and you whisper or mouth the words to yourself as if you were saying them. Nothing comes out of your mouth and nobody next to you could tell you were talking, but the words show up as text. There’s no wake word and no fixed list of commands, and you just say what you mean, silently.

The sound in the video is on the whole time. What you hear is the room, not me, because there’s nothing to hear. Every sentence that appears on screen was spoken with my mouth shut and my face still. If you watch the camera feed in the corner while the text comes in, there’s little to no visible movement to read either (and there’d be less with more practice!). From the outside it just looks like someone sitting quietly. Almost like telepathy!

So, what could this be used for? The obvious use is talking to a computer or an assistant in places where speaking out loud doesn’t work: an open office, a library, a train, a meeting you’re half-listening to. Past that, I can imagine it helping people who have lost their voice or can’t speak comfortably, private calls and dictation that nobody nearby can overhear, and communication in places too loud to be heard, like a factory floor or a flight deck. None of that is built or tested yet.

It was also cheap to make. The hardware is under $20 in materials, and the model was trained on about an hour of my own recordings plus synthetic data. Silent speech interfaces have been studied for a long time (Denby et al., 2010 survey the field), and the best-known systems set a high bar: MIT’s AlterEgo (Kapur et al., 2018) showed a wearable that recognizes internally articulated words, and Gaddy and Klein’s work on digitally voicing silent speech (2020) and its follow-up (2021) turned silently mouthed words into audio, trained on roughly 20 hours of data from a single speaker. Getting usable open-vocabulary results from one hour of real data and a sub-$20 build suggests the data and hardware bar is lower than it looks.

If this looks interesting to you, I’d be happy to chat more about it! You can reach out to me on X (@_asashepard_) or via email at asashepard@gmail.com.