Exploit comparable redundancy in raw-waveform speech encoders

Determine whether redundancy comparable to that exploited by stride-$k$ subsampling in Whisper can be exploited in raw-waveform encoders such as HuBERT and wav2vec~2.0.

Background

Stride-$k$ subsampling relies on overlapping windowed-frame preprocessing, particularly Whisper's log-mel frontend and convolutional stem, to ensure that retained tokens continue to cover the acoustic signal. Raw-waveform encoders such as HuBERT and wav2vec~2.0 construct representations directly from waveform inputs and therefore do not share this preprocessing structure. The paper reports that these encoders degrade sharply under stride-2 subsampling, but leaves unresolved whether some other form of comparable redundancy could nevertheless be identified and exploited in raw-waveform architectures.