Dynamic Encoder Size Based on Data-Driven Layer-wise Pruning for Speech Recognition

TL;DR

This paper introduces a data-driven, layer-wise pruning method to train a single supernet capable of producing multiple speech recognition models of different sizes, reducing redundancy and optimizing performance across various hardware constraints.

Contribution

It presents a novel dynamic encoder size approach that trains a supernet with shared parameters and automatically selects optimal subnets, streamlining model deployment for speech recognition.

Findings

Achieves comparable performance to individually trained models of various sizes.

Provides small performance improvements for the full-size supernet.

Reduces training and optimization efforts for multiple model sizes.

Abstract

Varying-size models are often required to deploy ASR systems under different hardware and/or application constraints such as memory and latency. To avoid redundant training and optimization efforts for individual models of different sizes, we present the dynamic encoder size approach, which jointly trains multiple performant models within one supernet from scratch. These subnets of various sizes are layer-wise pruned from the supernet, and thus, enjoy full parameter sharing. By combining score-based pruning with supernet training, we propose two novel methods, Simple-Top-k and Iterative-Zero-Out, to automatically select the best-performing subnets in a data-driven manner, avoiding resource-intensive search efforts. Our experiments using CTC on both Librispeech and TED-LIUM-v2 corpora show that our methods can achieve on-par performance as individually trained models of each size category. Also, our approach consistently brings small performance improvements for the full-size supernet.