Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text
A reported brain-to-text gain nearly disappears when the model is tested on signals with no brain information.
The paper says overlapping word-aligned windows leaked timing cues, letting the network infer words partly from speech duration. On synthetic non-brain signals, that setup reached 22.0% balanced accuracy, close to 22.3% on real recordings. Processing each window independently removed the shortcut and made repeated observations plus an LLM prior more useful. The authors report SimpleB2T reaching 36.6% word error rate with five observations per word on their perceived-speech benchmark. HF Daily Papers' note
The paper says overlapping word-aligned windows leaked timing cues, letting the network infer words partly from speech duration. On synthetic non-brain signals, that setup reached 22.0% balanced accuracy, close to 22.3% on real recordings. Processing each window independently removed the shortcut and made repeated observations plus an LLM prior more useful. The authors report SimpleB2T reaching 36.6% word error rate with five observations per word on their perceived-speech benchmark. HF Daily Papers' note
score 5