Two complementary tracks are offered: the Deep track targets within-subject word classification at scale, aiming for the best possible performance, while the Broad track targets cross-subject generalization, progressively reducing subject-specific fine-tuning data from ~40 to ~20 to ~10 minutes—the latter falling within a clinically feasible range. Winning submissions in the 2025 competition reached F1-macro scores of 95.6% and 73.6% on speech detection and phoneme classification tasks, respectively. This preprint has not been peer reviewed.
PNPL 2026 Challenges Teams to Adapt Speech Decoders on 10 Minutes of Data
Summary
A preprint posted to arXiv on September 3, 2026, sets out the 2026 PNPL competition, built on an extended LibriBrain100 dataset that adds 32 subjects at about 40 minutes each and roughly 80 more hours of within-subject data. The competition runs two tracks: a Deep track for within-subject word classification at scale, and a Broad track for cross-subject generalization that steps subject-specific fine-tuning data down from about 40 minutes to 20 and then 10, a clinically feasible range. Winning 2025 submissions reached F1-macro scores of 95.6% on speech detection and 73.6% on phoneme classification.
Why it matters
The number that matters here is 10 minutes. Non-invasive speech decoding is usually demonstrated on hours of data per subject, which is fine for a paper and unworkable in a clinic, so a track that forces adaptation inside a single appointment changes what counts as a competitive method. Competitions steer a field's effort more efficiently than individual papers do, and this one is pointed at transfer rather than peak accuracy.
Compiled by BCIwiki from public sources
Sources · 1
arxiv.org 2026-09-03