Conformer has been the ASR encoder to beat for years, but Branchformer and its follow-up E-Branchformer make a compelling case that sequential conv-plus-attention is not the only path. This review covers both papers back to back: Branchformer's parallel two-branch design (self-attention for global context, cgMLP for local dependencies) and E-Branchformer's enhanced merging strategy that pushes LibriSpeech test-clean and test-other WERs to 1.81% and 3.65% without external data.

The walkthrough unpacks why parallelizing local and global modeling gives you interpretability perks (the merging weights literally tell you how each layer balances the two), variable inference complexity from a single trained model, and better or matching accuracy versus Conformer. Note the errata called out at 13:40: it is a gating operation, not a ResNet connection. E-Branchformer builds on that by refining how the two branches fuse and stacking extra point-wise modules to squeeze out those state-of-the-art LibriSpeech WERs. If you are picking an encoder for your next ASR or SLU system, or just trying to understand where Transformer-family speech models are heading after Conformer, this side-by-side is a solid five-model-cards' worth of context in one sitting.