Open-content and zero-shot sEMG speech decoding

Develop reliable zero-shot and open-content surface EMG speech decoding that generalizes beyond the fixed 50-sentence corpus and does not require the closed-corpus sentence repetitions used in training.

Background

The primary experiments achieved short-calibration personalization only in a closed 50-sentence corpus. When the five evaluation sentences were excluded from all sEMG model-training data, decoding performance deteriorated sharply, showing that the reported personalization result does not establish generalization to unseen linguistic content. The paper therefore identifies reliable zero-shot and open-content decoding as unresolved.

References

These findings are specific to a cohort of healthy, speech-typical participants recorded with a standardized eight-channel montage and to a fixed 50-sentence corpus: decoding deteriorated sharply when the evaluation sentences were withheld from all sEMG model-training data, so reliable zero-shot and open-content decoding remain unresolved.

— Multi-Subject Pretraining Enables Short-Calibration Personalization for Closed-Corpus Surface EMG Speech Decoding  (2609.21288 - Le et al., 18 Sep 2026) in Conclusion