This paper was accepted on the Studying from Time Sequence for Well being workshop at NeurIPS 2025.
Each speech and sensor time collection knowledge encode data in each the time- and frequency- domains, like spectral powers and waveform shapelets. We present that speech basis fashions be taught representations that generalize past the speech area and obtain state-of-the-art efficiency on various time-series duties from wearable sensors. Probes educated on options extracted from HuBERT and wav2vec 2.0 outperform these extracted from self-supervised fashions educated straight on modality-specific datasets for temper classification, arrhythmia detection, and exercise classification duties. We discover that the convolutional function encoders of speech fashions are notably related for wearable sensor purposes. The proposed strategy enhances efficiency on knowledge scarce time-series duties utilizing easy probing strategies. This work takes a step towards creating generalized time-series fashions that unify speech and sensor modalities.
