Dynamic Encoder Size Based on Data-Driven Layer-wise Pruning for Speech Recognition
This addresses the problem of redundant training efforts for deploying ASR systems under varying constraints, but it is incremental as it builds on existing supernet and pruning techniques.
The paper tackles the need for varying-size ASR models under different hardware constraints by proposing a dynamic encoder size approach that jointly trains multiple performant models within one supernet, achieving on-par performance as individually trained models and small improvements for the full-size supernet.
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.