Effective Phase Encoding for End-To-End Speaker Verification

The widely used magnitude spectrum based features have shown their superiority in the field of speech processing. Incontrast, the importance of phase spectrum is always ignored.This is because the patterns hidden in phase cannot be intuitivelymodelled and interpreted, due to phase wrapping phenomenon.In this paper, we explore novel phase spectrum basedfeatures, named Learnable Group Delay (LearnGD), to captureuseful information in speech signals. Specifically, firstly, thenegative of the spectral derivative of the phase spectrum, calledgroup delay (GD), is used to unwrap the phase. Then, to suppressthe spiky nature of GD, which is caused by its roots closeto the unit circle in the Z domain, a carefully designed light convolutionalsmoothing layer is employed to reconstruct the GD.Finally, an exponential hyper-parameter is introduced to reconstructGD features to restore the spectrum range and generateLearnGD features. For performance evaluation, speaker verificationexperiments are conducted on the VoxCeleb2 corpus.Compared to the traditional acoustic feature derived from themagnitude spectrum, the proposed phase-based features reacha 27.8% relative improvement in terms of EER. Furthermore,experimental results on TIMIT phoneme recognition task alsodemonstrate the effectiveness of our proposed phase-based features.