A STATE-BASED NOISE-REDUCTION APPROACH FOR NONSTATIONARY ADDITIVE INTERFERENCE

In this paper we describe a technique that we developed for enhancing speech signals degraded by additive non-stationary noise. The performa nce of the technique is evaluated in the context of a speech recogniti on task on connected digits corrupted by different types of noise repr esentative of military environments. The algorithm is based upon spect ral amplitude estimation of the speech signal given state-dependent pa rametric speech and noise models. The spectral analysis is performed b y a resonator based frequency interpolation filterbank whose parameter s are selected according to the nature of the noise process. The model s are ergodic hidden Markov models (HMMs) with Gaussian multivariate d istributions trained on noise and speech samples.