Noise robust speaker verification with delta cepstrum normalization

This paper introduces a delta cepstrum normalization (DCN) technique for speaker verification under noisy conditions. Cepstral feature normalization techniques are widely used to mitigate spectral variations caused by various types of noise; however, little attention has been paid to normalizing delta features. A DCN technique that normalizes not only base features but also delta-features was recently proposed and showed high robustness in speech recognition and language identification. We introduce here DCN for a state-of-the-art speaker verification system that uses iVectors and probabilistic linear discriminant analysis. It is not obvious whether DCN is effective against speaker verification because DCN strongly transforms cepstral features and has possibility to distort the speaker-specific properties. We compared DCN with cepstral mean normalization (CMN), mean variance normalization (MVN), and histogram equalization (HEQ) using a NIST 2008 SRE dataset with various noise settings, and found that DCN is very effective even for speaker verification. DCN was especially effective under noisy conditions and achieved a maximum 18.5% relative error reduction compared to other competing methods. Combining verification scores from various feature normalization methods further improved the accuracy, and it achieved 9.1% and 16.4% relative error reduction under clean and noisy conditions, respectively.