Statistically Significant Duration-Independent-based Noise-Robust Speaker Verification

Abstract

A speaker verification system models individual speakers using different speech features to improve their robustness. However, redundant features degrade the system's performance. This paper presents Statistically Significant Duration-Independent Mel frequency Cepstral Coefficients (SSDI-MFCC) features with the Extreme Gradient Boost classifier for improving the noise-robustness of speaker models. Eight statistical descriptors are used to generate signal duration-independent features, and a statistically significant feature subset is obtained using a t-test. A redeveloped Librispeech database by adding noises from the AURORA database to simulate real-world test conditions for speaker verification is used for evaluation. The SSDI-MFCC is compared with Principal Component Analysis (PCA) and Genetic Algorithm (GA). The comparative results showed average equal error rate improvements by 4.93 % and 3.48 % with the SSDI-MFCC than GA-MFCC and PCA-MFCC in clean and noisy conditions, respectively. A significant reduction in verification time is observed using SSDI-MFCC than the complete feature set.

Keywords

• extreme gradient boost
• feature selection
• mel-frequency cepstral coefficients
• speaker verification