Search papers, labs, and topics across Lattice.
This study investigates the effectiveness of selective state-space models (Mamba) compared to recurrent decoders (GRU) in intracortical brain-to-text systems, focusing on the impact of output targets (phonetic vs. character). Using a controlled 2x2 experimental design on the Brain-to-Text'25 benchmark, the research finds that while the Mamba hybrid model is competitive, the GRU remains the superior choice, achieving lower phoneme and character error rates. Error analysis reveals distinct types of failures, highlighting the challenges in representation that affect performance based on the chosen output target.
GRU decoders outperform Mamba hybrids in brain-to-text systems, revealing critical insights into the interaction between output targets and model architecture.
State-of-the-art intracortical brain-to-text systems pair a neural-sequence phone decoder with an external language model. Two design axes remain underexplored: whether selective state-space models (Mamba) improve on recurrent decoders, and how the output target (phonetic vs.\ character) interacts with that choice. On the public Brain-to-Text'25 benchmark, we study a controlled 2x2 grid (GRU vs.\ hybrid Mamba decoder; phonetic vs.\ character targets) trained with a CTC objective under one reproducible protocol. The recurrent baseline remains strongest: the best phonetic GRU reaches 12.62\% PER and 21.19\% WER, while the best textual GRU after LM rescoring reaches 13.39\% CER and 26.28\% WER. The Mamba hybrid is competitive but does not surpass it. Ablations isolate architectural contributions, and error analysis shows representation-dependent failures: articulatory-like phoneme confusions vs.\ lexical and word-boundary errors.
