Performance improvement of speech recognizers in the presence of additive noise

Abstract

The problem of speech recognition in noisy environment has attracted the attention of many researchers. A speech recognition system designed to perform well under clean or low noise conditions often shows remarkable degradation in performance when background noise is present. In this dissertation work, we propose several methods to improve recognition accuracy under noisy conditions. The key problem of this dissertation work is the development of a speech recognition system that yields improved recognition performance at low noise conditions without causing degradation in performance at high noise conditions. The effectiveness of the proposed methods has been evaluated for additive white noise and additive colored noise using spectral distance comparison and speaker-independent isolated word recognition experiments. First, we propose a method of estimating autoregressive (AR) parameters in the presence of additive white noise. It is well known that when speech is modeled as an AR process and is contaminated by additive white noise, spectral zeros are introduced into speech spectrum and thus autoregressive moving-average (ARMA) modeling is more appropriate. The estimation problem for ARMA prcess is basically a nonlinear problem. Fortunately, however, it has been shown that the spectral poles of contaminated ARMA process are identical to those of original AR process. We assume that the introduced spectral zeros are closely ralated to the spectral poles. Based on this assumption, we estimate the AR parameters for clean speech through appropriate composite modeling of contaminated speech. We first filter contaminated speech by an all-pole filter which is the inverse of the estimated moving-average (MA) filter to cancel out the introduced spectral zeros, and then estimate the AR parameters from the filtered speech. This filtering and estimation procedures alternatively optimized through iterations. The amount of cancelling is adapted according to the estimate of sig...