Neural Noise

Real-time neural decoding of Mandarin Chinese for speech BCI

2 min read

  • Researchers achieved real-time decoding of full-spectrum Mandarin Chinese from neural signals using a brain-computer interface. 1
  • The work extends language-BCI beyond English to a tonal language and informs speech neuroprosthesis design. 1
  • The result was reported in Science/AAAS (November 2025). 1 1

Weekly enrichment (2026-07-20)

  • The underlying study is “Real-time decoding of full-spectrum Chinese using brain-computer interface” by Youkun Qian and colleagues (a Shanghai-based team), published in Science Advances on November 5, 2025 (vol. 11, issue 45, article eadz9968).23
  • The participant was a 43-year-old woman (an epilepsy patient) implanted with a 256-channel high-density, ultraconformal electrocorticography (ECoG) array, recorded across single-character and sentence reading tasks over 11 days.3
  • The system decoded a comprehensive set of 394 distinct Mandarin tonal syllables based purely on neural signals, reported as the first real-time BCI for a tonal monosyllabic language.23
  • Median single-character syllable identification accuracy was 71.2% (99% confidence interval 70.1–72.2%), with a long short-term memory (LSTM) architecture achieving the highest syllable decoding accuracy.2
  • Direct neural decoding alone yielded a character accuracy rate of 61.5%; after integrating a 3-gram Mandarin language model, real-time sentence decoding reached 73.1% character accuracy at a communication rate of 49.7 characters per minute.23
  • Analysis of the ECoG signals revealed distinct neural correlates for syllable and tone processing, and the tonally integrated system also let the participant control a robotic arm and digital avatar and interact with a large language model.3
  • Prior tonal-language work used a modularized multistream neural network that decoded lexical tones and base syllables via parallel streams from intracranial recordings during awake surgery, then synthesized speech—establishing the parallel tone/syllable decoding strategy this study builds on.4
  • For comparison, the English-language state of the art (Willett et al., Nature 2023) used intracortical microelectrode arrays in an ALS participant to reach 62 words per minute with a 9.1% word error rate on a 50-word vocabulary and 23.8% on a 125,000-word vocabulary, underscoring that tonal monosyllabic languages posed a distinct, previously unmet challenge.5

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE14dEFKZTgxTEFxX0I0X2R2ejk1LVQwSmc4WnpwMzNYUEtNRGVEYlNGUkhmSXpMaGsxVnVyTTFEOVkxYzBZVkVjX3Z6c3paSmQydGMteXp1dVhVYWhGVkdN?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.science.org/doi/10.1126/sciadv.adz9968 ↩ ↩2 ↩3 ↩4

  3. https://medicalxpress.com/news/2025-11-brain-interface-decodes-mandarin-neural.html ↩ ↩2 ↩3 ↩4 ↩5

  4. https://pmc.ncbi.nlm.nih.gov/articles/PMC10256166/ ↩

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

Graph View