Neural Noise

Speech prosthesis for clinical restoration (ALS, stroke)

3 min read

  • Rush University Medical Center has highlighted clinical speech restoration (e.g., for ALS and stroke) as part of “Finding a New Voice” (March 2025), with high relevance to neuroprosthetics and communication BCIs.1 1

Gardner updates

  • A streaming brain-to-voice neuroprosthesis decodes neural activity to naturalistic voice output for communication restoration, with strong implementation signal for near-term devices (Nature, tier-1). 2

  • WVU Rockefeller Neuroscience Institute was first in the U.S. to successfully test BCI technology for decoding speech and language. 3

  • Tracking single neurons in the human brain reveals new insight into language and other human-specific functions. 4

  • Human single-unit data underpins speech and language decoding for neuroprosthetics and is relevant to iEEG/speech BCI roadmaps. 4

  • ECoG decoding of words during sentence production reveals syntactic role encoding and structure-dependent temporal dynamics, directly advancing speech prosthesis and neural decoding. 5

  • A Stanford-led speech-enabling interface has demonstrated decoding for communication restoration in speech-impaired patients, with hope for restoring communication and tier-1 relevance for near-term translation. 6

Weekly enrichment (2026-07-20)

  • Willett et al. (2023, Nature) demonstrated an intracortical speech-to-text BCI in one ALS participant that decoded attempted speech at 62 words per minute (≈3.4× the prior record, approaching ~160 wpm natural conversation), with a 9.1% word error rate on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary — the first large-vocabulary demonstration.7
  • The rapidly-calibrating intracortical neuroprosthesis (Card et al., NEJM 2024) reached a 125,000-word vocabulary with useful function on the first day of use (25 days after implant), an initial Copy-Task word error rate of 9.8%, improving to 2.5% by session 15 and holding through session 84 (>8 months) at a 31.6 wpm self-paced rate.8
  • The UC Davis/BrainGate long-term study (Nature Medicine 2026) reported nearly two years and >3,800 cumulative hours of independent at-home use by ALS participant Casey Harrell: 183,060 sentences / 1,960,163 words at 56.1 wpm average, with 92.3% of sentences rated at least mostly correct and >99% word accuracy on a 125,000-word vocabulary in formal testing.9
  • That system used off-the-shelf Blackrock Neurotech microelectrode arrays and the BRAND real-time platform, and let the participant sustain full-time employment and independently operate his personal computer (speech BCI as keyboard, cursor BCI as mouse) with no researcher present.9
  • The instantaneous voice-synthesis neuroprosthesis (Wairagkar et al., Nature 2025) decoded from 256 microelectrodes in the ventral precentral gyrus of an ALS participant, synthesizing voice with closed-loop audio feedback at ~25 ms (one-fortieth of a second) latency and enabling real-time intonation control and short sung melodies.10
  • In that brain-to-voice study, listeners understood ~60% of synthesized words versus only ~4% from the participant’s residual speech, and neural-decoded voice was intelligible where microphone/stethoscope/IMU biosignals were not.10
  • The streaming brain-to-voice work (Littlejohn/Cho et al., Nature Neuroscience 2025) used high-density ECoG-style surface recordings in a stroke/anarthria participant (Ann Johnson) with an RNN-transducer decoding in 80-ms increments, personalized to her pre-injury voice, achieving ~47.5 wpm on the full vocabulary and 90.9 wpm on a 50-word set with ~1 s streaming latency (down from ~15 wpm in the team’s 2023 system).11
  • Clinical caveat: these are single-participant (n = 1) trials measured under controlled conditions with participant-specific calibration; communication rates reflect upper-bound lab performance, and replication across more participants and etiologies (e.g., stroke as well as ALS) remains the key unmet step before broad clinical deployment.12

Footnotes

  1. https://news.google.com/rss/articles/CBMiVkFVX3lxTE9laHVDVmhrYXB6Z3dWSnJGa1ZOXzNDdGtmZG9FZmpwZ2Q5YjNxQ3RXeHQxSWEwbERldWFaNEhFd2g3bnpWUi1yY3B3UDFpamk0SE03dUFn?oc=5 ↩ ↩2

  2. https://news.google.com/rss/articles/CBMiX0FVX3lxTE1yQkR6dTJXaDRsTnNoUjVQbGhUSVNzaF9UeFlzWEhZRHplRFV6b2tZM2NJSjVrMEI1eTZUTVhuNUtyeFFVWmxERzZaeG5XczQwZnpfM1E0N3lncTNIQmpv?oc=5 ↩

  3. https://news.google.com/rss/articles/CBMiwwJBVV95cUxPV0dpcDU2RUtyZXhycXZSYmRHdmlhWUM0aUFIQmFERkw1MzlqNTI5UnY4RkI5NEVscTIxOGJXdjlURU1nUzBnY2FsVHBuVnB0NTgyZGtsTldDQ3VDZmhxUzZ2cG5yUXBlOHV2SUE0eTZVNG10ZlEwWk00aG1ENmFWM2lqdWZtVjFuTFI5ZGl5LVRXYTYzakNtVEhtckF1RkpJQVBXOHA0czd5U3phSFpFbGRHdWJaUWJlRl93U0xLcHJLSVhXdzcxc1dSSG1PalRNNklweF9sa0JEUHF5RmRRYndsaGFMczRReU13WktVYkt2ay1aSVJDVWZZWVZqYkVNbFQyajRNUUpYbjhVajR1aWR4dkhDUWNRRVhjaEtEejQ4WklQcXBEWmlwcDJjOEE4WmR2b3R4OTgyTjcyd08tZ1RkOA?oc=5 ↩

  4. https://news.google.com/rss/articles/CBMi-AFBVV95cUxOVkM2UnlrTjdOQlc5VHhpOXowZHYtTWJaRDVzckYzSExZbnVVZENqOUt0QnpzQVBzQVQzNF9PcHlnSnpybTExVUt4WjNLbWYyRWFtXzZ2RUFpZkZ2WHpidG1lRlFiSUlNcGJ1VGdZY3VFSGZGeHZrbGlzNEJKNnhmaUF6TUI4d1ZPMWZ1c2lUbEpPUHZNcGU0Snd3Qk1aQ19hcTBMcGFqTjFsRnlIME9uMmk3dm94NDdySXMtT3owRlVBcjkxTkhEd0h6djIzX0JpVnBaWFFwY3VwSERZc3kyNE5vR2pJYVkzdEs0Z25xNHV6TlR3MVdZSQ?oc=5 ↩ ↩2

  5. https://news.google.com/rss/articles/CBMiX0FVX3lxTE5JNmU5WHlrdUJncmFTaUw2cmtmZVpaS25TdkpPODRoRjVkRW85SUNnZi11N2N0cktVN0FvdUhiS09ramt3SjUwWUNMcXNIOU5TN2tKQ09iYm9JTmU3SFp3?oc=5 ↩

  6. https://news.google.com/rss/articles/CBMiggFBVV95cUxPSHpXRkljdUFCdE1HT05qLWZYZzdJX2VLcXZCeHVvOWQtbm5qTEt0ZVBtcS12cVJSbkUxUmpobnYxUzlrQVdibVlOS2pONml1UHV5cmxhbTVNR0Jjd2s4NHBUSDdmN0hudnBNU0RJVmV0WFAtU0xEQkpkV3dGY1dFeVlB?oc=5 ↩

  7. https://www.nature.com/articles/s41586-023-06377-x ↩

  8. https://doi.org/10.1056/nejmoa2314132 ↩

  9. https://doi.org/10.1038/s41591-026-04414-6 ↩ ↩2

  10. https://www.nature.com/articles/s41586-025-09127-3 ↩ ↩2

  11. https://www.nature.com/articles/s41593-025-01905-6 ↩

  12. https://nhsjs.com/2026/brain-computer-interfaces-restoring-speech-in-paralysis-are-we-near-clinical-reality/ ↩

Graph View