A Joint Particle Filter and Multi-Step Linear Prediction Framework to Provide Enhanced Speech Features Prior to Automatic Recognition

Abstract

Automatic speech recognition, which works well on recordings captured with mid- or far-field microphones, is essential for a natural verbal communication between humans and machines. While a great deal of research effort has addressed one of the two distortions frequently encountered in mid- and far-field sound capture, namely non-stationary noise and reverberation, much less work has undertaken to jointly combat both kinds of distortions. In our view, however, this joint approach is essential in order to further reduce catastrophic effects of noise and reverberation that are encountered as soon as the microphone is more than a few centimeters from the speaker's mouth. We propose here to integrate an estimate of the reverberation obtained by multi-step linear prediction into a particle filter framework that tracks and removes non-stationary additive distortions. Evaluations on actual recordings with different speaker to microphone distances demonstrate that techniques combating either non-stationary noise or reverberation can be combined for good effect.