Word segmentation without known onsets

Solve the problem of segmenting words in non-invasive speech decoding without assuming that the onset time of each word is known.

Background

The paper evaluates word-aligned brain-to-text decoding, in which the temporal locations of perceived words are supplied to the decoder. This alignment assumption simplifies the task and enables extraction of fixed-length neural windows at known word onsets.

A practical non-invasive communication interface would need to identify word boundaries from neural recordings or another signal rather than receiving the onsets in advance. The authors explicitly state that this segmentation problem remains unsolved, distinguishing it from the word-aligned setting evaluated in the paper.

References

While we do not decode speech from paralysed patients, and have not yet solved the problem of segmenting words without knowing their onsets, SimpleB2T shows that in decoding perceived speech from non-invasive recordings, it is possible to achieve WERs much closer to that landmark invasive result.

— Removing Timing Shortcuts Improves Non-Invasive Brain-to-Text  (2609.40359 - Jayalath et al., 30 Sep 2026) in Section 4, “Towards useful non-invasive speech decoding”