Effective Phase Encoding for End-To-End Speaker Verification

conference paper

English

The widely used magnitude spectrum based features have shown their superiority in the field of speech processing. In contrast, the importance of phase spectrum is always ignored. This is because the patterns hidden in phase cannot be intuitively modelled and interpreted, due to phase wrapping phenomenon. In this paper, we explore novel phase spectrum based features, named Learnable Group Delay (LearnGD), to capture useful information in speech signals. Specifically, firstly, the negative of the spectral derivative of the phase spectrum, called group delay (GD), is used to unwrap the phase. Then, to suppress the spiky nature of GD, which is caused by its roots close to the unit circle in the Z domain, a carefully designed light convolutional smoothing layer is employed to reconstruct the GD. Finally, an exponential hyper-parameter is introduced to reconstruct GD features to restore the spectrum range and generate LearnGD features. For performance evaluation, speaker verification experiments are conducted on the VoxCeleb2 corpus. Compared to the traditional acoustic feature derived from the magnitude spectrum, the proposed phase-based features reach a 27.8% relative improvement in terms of EER. Furthermore, experimental results on TIMIT phoneme recognition task also demonstrate the effectiveness of our proposed phase-based features.

end-to-end speaker verification, phase information, group delay, on-the-fly

PENG, J.; QU, X.; GU, R.; WANG, J.; XIAO, J.; BURGET, L.; ČERNOCKÝ, J.

30. 8. 2021

International Speech Communication Association

Brno

1990-9772

Proceedings of Interspeech

2021

8

French Republic

2366

2370

5

https://www.isca-speech.org/archive/interspeech_2021/peng21c_interspeech.html

peng21c_interspeech.pdf