Decoding Intended Speech with an Intracortical Brain-Computer Interface in a Person with Longstanding Anarthria and Locked-In Syndrome | Jude et al.
Appreciation
--
Importance
--
Date Added
8.28.26
TLDR
In T17, 55b decodes phrases better than sentences, words, or phonemes; clusters speech by language category rather than mouth position; and decodes reading better than internal or attempted speech. This suggests 55b represents language content being held, not only speech being produced.
2 Cents
--
Tags
- 55b gets better at distinguishing stimuli as the language unit grows.
- Dorsal 55b had zero tuned electrodes for phonemes, more for words, a jump for phrases, and two more for sentences.
- The same pattern appeared in the single-word task, where 55b’s best-decoded words were the long ones.
- 55b separates speech sounds by category rather than mouth position, and it is loudest during reading.
- In a phoneme-only task, 6v clustered by tongue position. 55b’s top split was consonant versus vowel.
- With orofacial gestures mixed with phonemes, 6v again clustered by tongue position, while 55b’s top split was speech versus gesture.
- Decoding was strongest for reading, followed by internal speech, then attempted speech.
- 55b may encode meaning, but the evidence is preliminary.
- 55b confused two meaningless phrases despite their sharing no sounds, suggesting semantic coherence may matter more than articulation.
- This rests on only two nonsense stimuli and is not supported by the word-level control.
- 6v outperformed 55b on every articulatory task, but T17’s 55b arrays had much lower firing rates.
- 6v versus 55b accuracies were 50.4% versus 17.4% for phonemes, 91.5% versus 45.4% for orofacial gestures, and 51.1% versus 8.0% for single words.
- Kunz et al. also found very little spiking in T17’s 55b arrays and dropped them from decoding. T15’s 55b decoded well, suggesting an implant-specific limitation rather than a regional one.