Transferring Knowledge from a RNN to a DNN
This work addresses the challenge of making speech recognition models practical for embedded systems with limited computational capacity, though it is incremental as it builds on existing knowledge transfer methods.
The paper tackles the problem of deploying large neural networks on embedded systems by transferring knowledge from a state-of-the-art RNN to a small DNN, achieving a 3.93% WER on the WSJ eval92 task compared to a baseline of 4.54%, a 13% relative improvement.
Deep Neural Network (DNN) acoustic models have yielded many state-of-the-art results in Automatic Speech Recognition (ASR) tasks. More recently, Recurrent Neural Network (RNN) models have been shown to outperform DNNs counterparts. However, state-of-the-art DNN and RNN models tend to be impractical to deploy on embedded systems with limited computational capacity. Traditionally, the approach for embedded platforms is to either train a small DNN directly, or to train a small DNN that learns the output distribution of a large DNN. In this paper, we utilize a state-of-the-art RNN to transfer knowledge to small DNN. We use the RNN model to generate soft alignments and minimize the Kullback-Leibler divergence against the small DNN. The small DNN trained on the soft RNN alignments achieved a 3.93 WER on the Wall Street Journal (WSJ) eval92 task compared to a baseline 4.54 WER or more than 13% relative improvement.