BCIwiki (bciwiki.com) — A preprint uploaded to arXiv on 2026-10-16 introduces BEST (Brain-Encoding-Speech-Text), a two-stage framework for converting neural activity into text, to investigate whether speech brain-computer interfaces (BCIs) need larger models. It is reported here as a preprint that has not been peer reviewed. The authors separately state that it was accepted as a NeurIPS 2026 Main Track conference paper; these status statements require clarification.
Rather than relying on large generative models or phoneme-centric pipelines, BEST uses hierarchical intermediate representations and a compact speech-oriented decoder to map neural activity to text. Intermediate representations are information learned between the input signals and the text output. The authors describe the design as preserving essential structure while reducing computational cost and supporting settings with limited training data and computing resources.
On the Brain-to-Text '24 and '25 benchmarks, BEST achieved word error rates (WERs) of 6.19% and 1.62%, respectively, using only 3.85% of the parameters required by comparable methods. WER measures errors in the output relative to reference text, not communication success. The parameter ratio likewise does not directly specify runtime or energy consumption.
The authors draw three conclusions from their analysis: decoder scaling does not yield consistently increasing gains; cross-modal alignment matters more than model size; and overfitting degrades sequence-level performance. Representation analysis also reveals a consistent neural-audio-text alignment pathway. These are the authors' interpretations within this study, not evidence that all speech BCIs can dispense with large models.