Skipformer: A Skip-and-Recover Strategy for Efficient Speech Recognition

TL;DR

Skipformer tackles the quadratic cost of attention in Conformer-based ASR encoders by introducing a Skip-and-Recover strategy guided by an intermediate CTC. The encoder is split into two sub-blocks, with frames categorized into crucial, skipping, and ignoring groups; crucial frames pass through additional layers, skipping frames bypass layers, and ignoring frames are discarded, with recovered outputs reassembled in temporal order. An intermediate loss ties the CTC and AED objectives to the frame-splitting mechanism, and five data-splitting modes are explored to balance accuracy and efficiency, with Mode 2 typically offering the best trade-off. Experiments on AIShell-1 and Librispeech show substantial reductions in input length (up to ~31x) and faster inference, while maintaining or improving accuracy; the method is open-sourced for broader adoption.

Abstract

Conformer-based attention models have become the de facto backbone model for Automatic Speech Recognition tasks. A blank symbol is usually introduced to align the input and output sequences for CTC or RNN-T models. Unfortunately, the long input length overloads computational budget and memory consumption quadratically by attention mechanism. In this work, we propose a "Skip-and-Recover" Conformer architecture, named Skipformer, to squeeze sequence input length dynamically and inhomogeneously. Skipformer uses an intermediate CTC output as criteria to split frames into three groups: crucial, skipping and ignoring. The crucial group feeds into next conformer blocks and its output joint with skipping group by original temporal order as the final encoder output. Experiments show that our model reduces the input sequence length by 31 times on Aishell-1 and 22 times on Librispeech corpus. Meanwhile, the model can achieve better recognition accuracy and faster inference speed than recent baseline models. Our code is open-sourced and available online.

Figures (1)

• Figure 1: Overview of the model architecture. a) model architecture; b) split mode.