Prototype Tongue-Reading System Decodes Speech With 15.6% Error Using 50 Hours of Training
Updated
Updated · Hackaday · Jul 12
Prototype Tongue-Reading System Decodes Speech With 15.6% Error Using 50 Hours of Training
1 articles · Updated · Hackaday · Jul 12
Summary
A prototype speech system converted silent tongue movements into decoded words with a 15.6% error rate after training on about 50 hours of ultrasound data.
Ultrasound video from a probe under the chin captures tongue motion instead of sound, letting users “speak” privately and keeping the system unaffected by noisy environments.
The model also generalized across different speakers with similar accents, suggesting accuracy could improve with larger datasets and more training.
Its main limitation is hardware: the current off-the-shelf probe must be held under the chin, though a smaller adhesive-patch design could make the approach wearable.