On the Compression of Recurrent Neural Networks with an Application to LVCSR acoustic modeling for Embedded Speech Recognition

TL;DR

This paper presents a method for compressing recurrent neural networks, specifically LSTM acoustic models, to enable efficient speech recognition on mobile devices without significant accuracy loss.

Contribution

It introduces a joint compression technique for recurrent and non-recurrent weights, significantly reducing model size while maintaining performance.

Findings

LSTM model size reduced to one-third of original

Negligible accuracy loss after compression

Applicable to embedded speech recognition systems

Abstract

We study the problem of compressing recurrent neural networks (RNNs). In particular, we focus on the compression of RNN acoustic models, which are motivated by the goal of building compact and accurate speech recognition systems which can be run efficiently on mobile devices. In this work, we present a technique for general recurrent model compression that jointly compresses both recurrent and non-recurrent inter-layer weight matrices. We find that the proposed technique allows us to reduce the size of our Long Short-Term Memory (LSTM) acoustic model to a third of its original size with negligible loss in accuracy.