Consolidating Product Spectrum and Gammatone filterbank for robust speaker verification under noisy conditions

Abstract

Motivated by recent advances in speech and audio processing community reporting that incorporating the phase information can improve further the performance of state-of-the-art phase-independent features. We propose in this paper a modification in the extraction pipeline of the Mel-frequency Product Spectrum Cepstral Coefficients MFPSCC which warp the product spectrum with a Mel- Scale filterbank. The main novelty of this work resides in incorporating a Gammatone filterbank as a substitute of the Mel filterbank to reinforce the robustness of whole speaker verification system in noisy conditions. The proposed feature is dubbed the Gammatone Product-Spectrum Cepstral coefficients GPSCC. Experimental results are undertaken on the TIMIT corpus corrupted by different stationary and non-stationary noises using the GMMUBM de facto standard for speaker verification. Performance evaluations demonstrate drastic reduction in Equal Error Rates when using GPSCC compared to other related features and this gain in performance is more pronounced at low signal to noise ratios.