Getting started
This challenge encourages broad participation and aims to advance the societal impact of speech decoding technologies. Everyone is welcome — no neuroscience background or specialised hardware required. Free GPU access is available through Google Colab.
Tracks and divisions
The competition uses the LibriBrain100 dataset and has two tracks, both evaluated on word classification (top-10 balanced accuracy over a vocabulary of 50 words):
| Track 1 — Deep | Track 2 — Broad | |
|---|---|---|
| Goal | Word classification on a single, deeply-sampled subject (subj0) | Word classification across 32 held-out subjects (subj1–subj32) |
| Training data | Extensive recordings from subj0 across audiobooks, phonetically balanced speech corpora (TIMIT, MOCHA-TIMIT), and narrative podcasts | Varying amounts of labelled data per subject (see below) |
Track 2 data groups. To study how performance scales with the amount of fine-tuning data, subjects in Track 2 are split into groups:
| Group | Subjects | Labelled data per subject |
|---|---|---|
| 100% | 12 | ~40 min of audiobook listening |
| 50% | 10 | ~20 min of audiobook listening |
| 25% | 10 | ~10 min of audiobook listening |
Submissions and leaderboards
- You may submit to either or both tracks. Your progress will appear on the relevant leaderboard.
- Final rankings are determined on independent holdout data recorded with stimuli that are entirely different from anything released to participants.
- Submissions must beat the reference baselines to be eligible for prizes. Baseline values will be published ahead of the submission window.
Prizes
- The top 3 confirmed teams per track win prizes, provided they beat the baselines.
- Each team may only win prize money once across the entire competition. If the same team places in the top 3 on multiple leaderboards, they will still be listed on all of them, but the prize for each additional placement goes to the next-ranked team.
- In the unlikely event of a tie, the prize will be split.
See the Prizes page for details.
Verification
After the submission deadline, all top-ranking teams will be asked to provide runnable model checkpoints. We may also ask specific teams to provide their training code so we can verify no cheating took place.
The top 3 teams per track whose submissions are confirmed through this process will be declared the winners.
Open science
We strongly encourage all participants to share their code openly. We also welcome pull requests to the PNPL library — for example, adding data loaders for new datasets to accelerate the community.
Communication
We have set up a Discord server, which already has an active community of researchers working on neural speech decoding. Join here: Discord invite link.