Intracortical Brain-to-Text: GRU Decoders Outperform Mamba on Phoneme and Character Targets

A recent investigation into the Brain-to-Text '25 benchmark evaluates recurrent (GRU) and selective state-space (Mamba) decoders for intracortical brain-to-text applications, focusing on phoneme and character targets. The top-performing phonetic GRU records a 12.62% PER and a 21.19% WER, while the leading textual GRU, following language model rescoring, achieves a 13.39% CER and a 26.28% WER. Although the Mamba hybrid shows competitive results, it falls short of the recurrent baseline. An analysis of errors indicates that failures are dependent on representation, highlighting confusions in articulatory-like phonemes compared to lexical and word-boundary mistakes.