Key Takeaways
Aleph, a brain-interface lab, said on July 7 that it used ultrasound held under the chin to read silent speech by imaging the tongue, with no sound spoken.
Trained on about 50 hours of data over roughly a month, the model handled an open vocabulary and, by Aleph's account, beat every existing silent speech method except camera-based lip-reading.
The model generalized to people it had never recorded, as long as they had an American accent, an early result shared on X but not yet peer reviewed.
Aleph, a brain-interface lab, said on July 7 that it built a system that reads silent speech with ultrasound, letting a person talk to a computer without making a sound. The claim, posted in a thread on X that drew almost 200,000 views, describes a probe held under the chin that images the tongue as it moves.
The lab said the idea started in December, when it put a transducer under the chin and found it could see the tongue. A first test used a one-hour dataset of 20 phrases and a simple classifier that reached 95 percent accuracy, enough to suggest silent speech could be approached with ultrasound at all.
Last month Aleph ran a larger project, collecting about 50 hours of data and training a model to handle any word rather than a fixed bank of sentences. By its own account, that model beat every available silent speech method on open vocabulary except lip-reading, which needs a camera pointed at the face.
The part the lab called surprising was that the model already worked on people it had never recorded, as long as they had an American accent. Aleph described the system on X as "an earlier version of telepathy."
Silent speech interfaces are not new. Research groups have mapped the tongue with ultrasound for years, and the work matters most for people who cannot vocalize, including those living with ALS or recovering from a laryngectomy. What is new is the jump to open vocabulary from a single month of data.
The results have not been peer reviewed, and the accent limit shows how far the generalization still has to travel. The same lab captured the first super-resolution ultrasound images of a living human brain in June, and this fits the same push to read the body without cutting into it, the goal behind Neuralink threading its implant through the intact dura. Worth watching whether the vocabulary and the accents both widen.
People Also Ask
How does Aleph's ultrasound silent speech work?
A probe held under the chin sends ultrasound into the mouth and images the tongue as it moves, and a model maps those tongue movements to words, with no sound spoken aloud.
What is silent speech technology used for?
Silent speech interfaces let people communicate without vocalizing, which matters most for people who have lost their voice to ALS, stroke, or a laryngectomy, and for private, hands-free computer control.
Did Aleph's model work on new people?
Aleph said the model generalized to people it had never recorded, as long as they spoke with an American accent, though the results have not been peer reviewed.
How much data did Aleph use to train the model?
Aleph said it trained on about 50 hours of data collected over roughly a month, after an initial one-hour test of 20 phrases reached 95 percent accuracy.
Sources: Aleph (X thread and blog), arXiv (SottoVoce silent speech research).