Cepstral gain normalization for noise robust speech recognition

The paper describes a robust speech recognition technique which normalizes cepstral gains in order to remove effects of additive noise. We assume that the effects can be expressed by an approximate model which consists of gain and DC components in log-spectrum. Accordingly, we propose cepstral gain normalization (CGN) which normalizes the gains by means of calculating maximum and minimum values of cepstral coefficients in speech frames. The proposed method can extract noise robust features without a priori knowledge and environmental adaptation because it is applied to both training and testing data. We have evaluated recognition performance under noisy environments using the Noisex-92 database and a 100 Japanese city names task. The CGN provides improvements of recognition accuracy at various SNRs compared with combinations of conventional methods.