Significance of excitation source sequence information for Speaker Verification

Abstract

Automatic speaker verification (ASV) system is used to verify the speaker’s identity. Over the years, many spectral and excitation source features have successfully contributed to improving the performance of the ASV. In previous works, it has been shown that there is complementary information in the excitation sources. Due to the variation in the physiological structure across speakers, the generated excitation varies from speaker to speaker. In this work, we have performed a comparative study to analyze the extent of speaker signatures in each excitation source. The excitation sources over which comparison is made are epoch, weighted epoch, Hilbert envelope of linear prediction (LP) residual, residual phase, and Gaussian noise. The comparison with the Gaussian noise sequence shows that excitation sources considered noisy sequences have better speaker-specific information than the noise. Pearson correlation coefficient, structural similarity index measure (SSIM), and equal error rate (EER) were used as the metrics to evaluate the performances. The ASV performance is tested on TIMIT and IITG-MV datasets.