Robust speaker verification with a two classifier format and feature enhancement

In the presence of environmental noise, speaker verification systems inevitably see a decrease in performance. This paper proposes the (1) use of two parallel classifiers, (2) feature enhancement based on blind signal-to-noise ratio (SNR) estimation and (3) fusion, to improve the performance of speaker verification systems. The two classifiers are based on Gaussian mixture models and the partial least-squares technique. Speech corrupted by additive noise at SNRs from 0 to 30 dB are used for authentication. A two-way analysis of variance validates the performance gain offered by the methods used. The outputs of the classifiers are fused together in different ways. The fusion method where the scores of the classifiers are added together is found to be the best method again using statistical analysis.