A transformer block with a convolution module inside - attention for global context, convs for local acoustic detail.
The Conformer sandwiches a depthwise-convolution module between attention and feed-forward layers (in a macaron FFN-attention-conv-FFN pattern), usually with relative position encoding in the attention. Convolutions capture local spectral patterns that pure attention handles poorly, which made Conformers the dominant speech-encoder architecture (NVIDIA's FastConformer here).
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 →