Strided convolutions shrink the audio sequence (often 4–8×) before the expensive transformer layers run.
Spectrogram frames arrive at ~100 per second - far denser than needed. A small stack of strided convolutions downsamples time resolution several-fold and lifts features into the model width, cutting transformer cost quadratically. Practically every ASR encoder (Conformer, Whisper, Parakeet) starts this way.
Open any of these on hfviewer to find this block in the interactive architecture graph.
hfviewer renders the full architecture of 2,300+ Hugging Face models as interactive graphs - hover any block to see what it does, with this model's real numbers.
Browse all model graphs →