Zipformer: A faster and better encoder for automatic speech recognition

TL;DR

Zipformer is an improved, faster, and more memory-efficient encoder for automatic speech recognition that outperforms existing models through novel architecture modifications and a new optimizer.

Contribution

The paper introduces Zipformer, a novel encoder architecture with innovative modeling techniques and a new optimizer, achieving superior ASR performance and efficiency.

Findings

Outperforms state-of-the-art ASR models on multiple datasets

Faster convergence with the new ScaledAdam optimizer

More memory-efficient and better-performing encoder architecture

Abstract

The Conformer has become the most popular encoder model for automatic speech recognition (ASR). It adds convolution modules to a transformer to learn both local and global dependencies. In this work we describe a faster, more memory-efficient, and better-performing transformer, called Zipformer. Modeling changes include: 1) a U-Net-like encoder structure where middle stacks operate at lower frame rates; 2) reorganized block structure with more modules, within which we re-use attention weights for efficiency; 3) a modified form of LayerNorm called BiasNorm allows us to retain some length information; 4) new activation functions SwooshR and SwooshL work better than Swish. We also propose a new optimizer, called ScaledAdam, which scales the update by each tensor's current scale to keep the relative change about the same, and also explictly learns the parameter scale. It achieves faster convergence and better performance than Adam. Extensive experiments on LibriSpeech, Aishell-1, and WenetSpeech datasets demonstrate the effectiveness of our proposed Zipformer over other state-of-the-art ASR models. Our code is publicly available at https://github.com/k2-fsa/icefall.