A state-based noise reduction approach for non-stationary additive interference

Abstract

In this paper we describe a technique that we developed for enhancing speech signals degraded by additive non-stationary noise. The performance of the technique is evaluated in the context of a speech recognition task on connected digits corrupted by different types of noise representative of military environments. The algorithm is based upon spectral amplitude estimation of the speech signal given state-dependent parametric speech and noise models. The spectral analysis is performed by a resonator based frequency interpolation filterbank whose parameters are selected according to the nature of the noise process. The models are ergodic hidden Markov models (HMMs) with Gaussian multivariate distributions trained on noise and speech samples.