The IBM 2015 English Conversational Telephone Speech Recognition System

TL;DR

This paper details improvements to IBM's English conversational telephone speech recognition system, achieving a significant reduction in word error rate through advanced neural network techniques and language model rescoring.

Contribution

Introduction of novel neural network architectures and training strategies, combined with sophisticated language model rescoring, to enhance speech recognition accuracy.

Findings

8.0% word error rate on Switchboard test set

23% relative improvement over previous best

Effective use of maxout networks and joint modeling techniques

Abstract

We describe the latest improvements to the IBM English conversational telephone speech recognition system. Some of the techniques that were found beneficial are: maxout networks with annealed dropout rates; networks with a very large number of outputs trained on 2000 hours of data; joint modeling of partially unfolded recurrent neural networks and convolutional nets by combining the bottleneck and output layers and retraining the resulting model; and lastly, sophisticated language model rescoring with exponential and neural network LMs. These techniques result in an 8.0% word error rate on the Switchboard part of the Hub5-2000 evaluation test set which is 23% relative better than our previous best published result.