Within-Hour Signal Drift and Blocked Recording: What a Labelled Continuous Recording of Silent Articulation Shows
A continuous recording 59.9 minutes long, within which all 199 silent-articulation trials of one day are labelled, makes it possible to test what segmented trials cannot show: how the signal changes over a session and what a cross-validation actually measures. When words are recorded in blocks of 20 consecutive trials, recording time and the spoken word become coupled, and shuffled cross-validation partly measures recognition of time rather than of the word. On these data, shuffled validation gives 72% on five classes, whereas validation that trains on one set of blocks and tests on the other gives 41% and 21% in the two directions, against a chance level of 20%. The first block of a word is distinguished from the second block of the same word with 85–100% accuracy, so the session fingerprint within a single day is stronger than the differences between words. Separately, trials of the "silence" class are shown not to be rest: their amplitude is 1.5 times that of the inter-trial intervals. A protocol requirement follows: words must be presented interleaved in random order, otherwise word and time cannot be separated by any processing. Russian and English versions attached.
Authors
- Fedotova Ann (ORCID: https://orcid.org/0009-0008-7300-3460)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24279767
- Primary Topic
- Speech Recognition and Synthesis
- Type
- preprint