Developer Rebuilds Speech BCI Decoder, Cuts Word Error Rate to 23.5%
10/03/2026 — 10/03, 15:12·1 sources·1 reports
Story overview
On October 3, 2026, a developer published a post on DEV Community describing how they rebuilt a speech brain-computer interface decoder. The starting point, as the author frames it, is that people who can no longer speak can still try to speak, and the motor cortex activity produced by that attempt can be decoded into text.
The author rebuilt the decoder from the neural data released with Willett et al.'s speech BCI paper in Nature 2023. According to the post, the model trained on CPUs for about five hours. Over the course of four changes, word error rate on a test set of 880 sentences dropped from 50.0% to 23.5%. After the final change, the rate measured 23.0% on a held-out half of the data.
The author also states that the work is an offline reanalysis of already-recorded data, not a new clinical trial. That is where matters currently stand, and the single report contains no further developments.
AI-generated from 1 reports · updated 1 hour ago
Latest turnA developer rebuilt a brain-to-text decoder from the neural data released with Willett et al. (Nature 2023), training on CPUs for about five hours. Across four changes, word error rate on 880 test sentences fell from 50.0% to 23.5%, and reached 23.0% on a held-out half. It is an offline reanalysis of recorded data, not a new clinical trial.
Reports on this story headlines open the original
A developer rebuilt a brain-to-text decoder from the neural data released with Willett et al. (Nature 2023), training on CPUs for about five hours. Across four changes, word error rate on 880 test sentences fell from 50.0% to 23.5%, and reached 23.0% on a held-out half. It is an offline reanalysis of recorded data, not a new clinical trial.
DEV Community · AIAI score 62
Other stories people are talking about
- 508NVIDIA launches 64GB DGX Spark desktop AI computer at $4,9998 sources
- 494Apple tightens macOS Full Disk Access over AI agent risks7 sources
- 367Meta open-sources Muse Gadgets firmware and SDK5 sources
- 246SurgeMicrosoft AI releases MAI-Transcribe-2-Streaming real-time transcription model4 sources
- 215SurgeAmazon Weighs Moving $8 Billion of Nvidia Chips into a Financing Vehicle3 sources
- 208Anthropic launches Claude Frontier Academy with $100M3 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion