Statistical feature compensation and normalization for speech recognition in noisy environments

Abstract

The statistical mismatch of the speech feature between the training environment and the testing environment results in performance degradation of cepstrum based speech recognition system. In this dissertation work, we propose a new feature compensation method and two novel feature normalization algorithms based on statistical methods. When speech signals are contaminated by additive noise, the statistical properties of a speech feature vector vary according to the types of noise and signal-to-noise ratio (SNR) levels. The mutivaRiate-gAussian-based cepsTral normaliZation (RATZ) is one of the best known EM-based feature compensation method. However, the noisy model of the RATZ only represents the mean shift of feature vector. We propose a new noisy model for the RATZ which represents the variance deviation, as well as the mean shift. For the feature normalization algorithm, conventional methods only normalize the mean and/or variance of the cepstrum features. However, deviations of higher order moments also exist in noisy speech features. In order to fully normalize the variations of the statistical properties under noisy conditions, all the moments or the probability density functions (pdf) must be normalized. As the first step to full normalization, we propose cepstrum third-order normalization (CTN) method, which normalizes the third-order moment of the cepstrum as well as mean and variance. Moreover, we propose cepstrum pdf normalization (CPN) method, which fully normalizes the statistical properties. In order to consider various densities, the generalized Gaussian distribution (GGD) is used as the target pdf. A table lookup method is also used to alleviate the computational load of the CPN. From the speaker-independent word recognition experiments, we show that the propose methods give improved performance compared with that of the conventional methods, especially in heavy noise environments.