Constrained Spectrum Normalization for Robust Speech Recognition in Noise

This paper presents a new approach to robust speech recognition in noise based on spectral subtraction. A conventional spectral subtraction technique leads to nonlinear distortions of the normalized speech signals and resulting degradation of speech recognition accuracy. A new method is proposed to constrain spectral subtraction by imposing upper bounds on the estimates of the noise spectra. Two speech databases collected in moving cars were used in speech recognition experiments. A set of cross-database recognition experiments revealed that this technique is capable of improving robustness of a speech recognition system. When HMMs trained on the data from one database were used to recognize the data from another database, relative string error rate reduction of 20% to 45% was obtained by using the proposed method.