269 Hours of EEG Data: Non-Invasive Speech Decoding Hits 61.3% Accuracy
kaixhin · x · 2026-08-10
Araya's team published a new research chapter on non-invasive EEG-based speech decoding. The researchers collected 269 hours of EEG data during vocalized speech in an open vocabulary setting and trained a self-supervised model.
Key findings include:
- Scaling Effects: The model demonstrated continuous improvement with more data while effectively mitigating muscle artifacts, achieving a top-1 accuracy of 61.3% in classifying 512 speech segments.
- Temporal Generalization: Tested seven months after initial training, the model maintained its accuracy requiring only 30 minutes of calibration data.
- Participant Transferability: Models pre-trained on large-scale datasets showed a 19% improvement in top-1 accuracy when fine-tuned with new participant data, compared to models trained from scratch.
More from Research
- Nature: Artificial Neural Manifolds via Mott Memristors Enable Efficient Prediction — AnnaCiaunica · 2026-08-10
- Economist Chad Jones Publishes Paper on AI and Our Economic Future in JEP — davilagrau · 2026-08-10
- PIXIO Report: Doubles Open Video Model Speed on Single 96GB GPU — tsi_org · 2026-08-10
- Why Do We Say 'Thank You' to LLMs? Paper Introduces Embodied Hijack Hypothesis — MacrinePhD · 2026-08-10
- CAS Introduces CtrlBench-Rec: Multi-Agent Framework for Recommender Controllability — jiqizhixin · 2026-08-10
- Google DeepMind Releases 'How To Scale Your Model' Systems Guide — tetsuoai · 2026-08-10