Neural NoiseSpeech bci-and-neural-decoding aim to restore communication by decoding speech-related neural signals directly from the brain. These systems target individuals who have lost the ability to speak due to conditions such as amyotrophic lateral sclerosis, brainstem stroke, or severe spinal cord injury, offering the potential for naturalistic, high-bandwidth communication that surpasses letter-by-letter BCI spelling approaches.
Early speech BCI research established that neural signals associated with speech production, including articulatory movements and phonemic representations, could be recorded from speech motor cortex and decoded into text or synthesized speech. Both intracortical microelectrode arrays and high-density ecog grids have been used to capture the distributed neural activity underlying speech, with decoding vocabulary growing from individual phonemes to sentences of increasing complexity.
Key technical challenges include achieving real-time decoding speeds that match natural speech rates, handling the variability of attempted speech in paralyzed users who cannot produce overt movements, and building robust language models that integrate neural and linguistic information. The convergence of high-channel-count neural recording, deep learning sequence models, and large language models is accelerating progress toward clinical speech BCI systems.
Gardner updates
-
The Transmitter argues chronic speech BCIs let researchers study how speech is produced in everyday life, not only controlled lab tasks, linking neural decoding to speech-prosthesis agendas. 1
-
Neuroscience News coverage reports an implanted AI speech neuroprosthesis in advanced ALS with about two years of real-time use (~2.7 million words), including intonation modulation and singing. 2
-
A bioRxiv preprint proposes MEG-informed navigated rTMS speech cortical mapping that uses individual MEG speech-production maps to improve noninvasive localization of speech networks for neurosurgery and speech-BCI targeting. 3
-
A Journal of Neural Engineering EEG silent-BCI study decodes imagined Chinese speech with a capsule network and bidirectional knowledge transfer for hierarchical multi-label phoneme classification. 4