Build log
Can a ring understand a mumble?
A ring hears you from your hand, not from your mouth, and often while you move. Before building one, we tested how well speech recognition copes with that, in six languages.
In short. Through a simulated ring microphone, a large open speech-recognition model understood a quiet room well in all six languages we tried, with about half the errors of a smaller model at the same speed. Tuning a model on our simulated sound did not carry to speech it had not heard, and later attempts with more realistic sound have not been good enough to use either.
The question
Speech recognition is usually tested with a microphone in front of the speaker. A ring is different: it sits on your finger, it moves, and it hears your voice from wherever your hand happens to be. If the ring cannot understand you, nothing else it does matters.
How we tested
There is no ring yet, so we built its sound path in software: the distance from the mouth, the microphone behind its waterproof membrane, a moving hand, background noise, and the compression the ring uses to send your voice. Then we ran recorded speech through it:
- read speech in English, Turkish, German, French, Spanish and Italian (the FLEURS corpus);
- English audiobooks (LibriSpeech) and European Parliament speeches (VoxPopuli), which no tuning ever touched;
- quiet rooms, noise, and mumbled speech.
Every result below comes from that simulated path, not from a ring. Each one stands as an assumption until a prototype records real voices.
What we found
The larger model made about half as many mistakes as the smaller one: in a quiet room, in Turkish as in the other languages, and when people mumbled.
It won every pairing of language and condition we tried, at about the same speed on our graphics card. As released, it is the candidate for our cloud.
What did not work
We tried to make the small model better by training it on speech passed through the simulated ring. The first version, trained on English, understood English better in noise and five other languages worse. The second, trained on all six languages, improved most tests on the corpus it trained on, and showed no clear gain on speeches it had not heard before. Applied to the larger model, it made things worse on unseen speech.
The lesson: a simulation teaches a model the simulation. We tried again with richer scenes and real meeting recordings, under rules written down before each run. None of those versions was good enough to use either, and one learned to invent words: what our ring and robot learn. Recordings from the ring itself, judged on speech no training has touched, will decide.
When it is not sure
A mumble is still hard for every model we tried, and much harder in some languages than in others. So the ring does not guess. When the recogniser is unsure, it buzzes and shows SAY AGAIN. And if you spoke too far from it, the ring can tell before sending: when your voice does not stand far enough above the noise, it shows HOLD ME CLOSER. In simulation, that cue caught many of the sentences that would have come back badly garbled, and rarely interrupted one that was fine.
What is next
A development board with the ring’s radio and microphone, real voices in real rooms, and the same tests again. We will write here what we find.
Why speech recognition errs more for some voices than others, and what helps: voice assistants and accents.