Silent Articulation from Neck and Submental EMG: the Temporal Fingerprint as the Main Limit of Wearable Recognition
We describe a prototype of a wearable silent speech interface in the form of a necklace. The prototype records surface electromyography (EMG) from the neck and the submental area and recognises words mouthed without sound. The dataset contains 319 trials from one speaker over three days: 8 words and a "silence" class. Within a recording session the model distinguishes 5 words with 76% accuracy (chance 20%) and 3 words with 88% accuracy. We show that the signal carries a fingerprint of the recording time. The day of recording is identified from a single "silence" trial in 100% of cases. Within a day, two series of the same word recorded 6–19 minutes apart are told apart in 85–100% of cases. When tested on a series recorded at another time, accuracy drops to 21–41% on 5 words, and to chance on another day. Conclusion: the main limit of wearable recognition is signal drift over time. It is addressed by calibration when the device is put on and by recording words in random order. Russian and English versions attached. The author is the founder of VOIZEX. Data processing and draft text were prepared with the assistance of Claude (Anthropic); the author reviewed the text and is responsible for it.
Authors
- Fedotova Ann (ORCID: https://orcid.org/0009-0008-7300-3460)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24279761
- Primary Topic
- Muscle activation and electromyography studies
- Type
- preprint