Defeating reverberation: Advanced dereverberation and recognition techniques for hands-free speech recognition

Abstract

Automatic speech recognition is being used successfully in more and more products. However, current recognition systems usually require the use of close-talking microphones. This constraint limits the deployment of speech recognition for new applications. In hands-free situations, noise and reverberation cause a severe degradation of the recognition performance. The problem of noise robustness has attracted a great deal of attention and practical solutions have been proposed and evaluated with common benchmarks. In contrast, reverberation has long been considered an unsolvable problem. Recently, significant progress has been made in the field of reverberant speech recognition and this progress has been evaluated with the REVERB challenge 2014. In this paper, we describe the reverberant speech recognition system we proposed for the REVERB challenge that exhibited high recognition performance even under severe reverberation conditions. We compare our system with other proposed approaches to suggest potential future research directions in the field.