Phoneme vs. Character Targets and Mamba Decoders for Intracortical Brain-to-Text

This paper evaluates two underexplored design choices in intracortical brain-to-text: recurrent versus selective state-space decoders, and phoneme versus character targets. Using a controlled 2x2 study on the public Brain-to-Text ’25 benchmark, all systems are trained with CTC under one reproducible protocol. The strongest phoneme GRU achieves 12.62% phoneme error rate (PER) and 21.19% word error rate (WER). The best character GRU, after language-model rescoring, reaches 13.39% character error rate (CER) and 26.28% WER. The hybrid Mamba decoder is competitive but does not surpass the recurrent baseline.