Spectral subtraction and missing feature modeling for speaker verification

This paper addresses the problem of robust text-independent speaker verification when some of the features for the target signal are heavily masked by noise. In the framework of Gaussian mixture models (GMMs), a new approach based on the spectral subtraction technique and the statistical missing feature compensation is presented. The identity of spectral features missing due to noise masking is provided by the spectral subtraction algorithm. Consequently, the statistical missing feature compensation dynamically modifies the probability computations performed in GMM recognizers. The proposed algorithm uses a variation of the generalized spectral subtraction and incorporates in it a criterion based on masking properties of the human auditory system. The originality of the algorithm resides in the fact that instead of using fixed parameters for the noise reduction and missing feature compensation, the noise masking threshold is used to control the enhancement and model compensation processes adaptively, frame-by-frame, hence helping to find the best tradeoff.