ICSpk: Interpretable Complex Speaker Embedding Extractor from Raw Waveform

Recently, extracting speaker embedding directly from raw waveformhas drawn increasing attention in the field of speaker verification.Parametric real-valued filters in the first convolutionallayer are learned to transform the waveform into time-frequencyrepresentations. However, these methods only focus on themagnitude spectrum and the poor interpretability of the learnedfilters limits the performance. In this paper, we propose a complexspeaker embedding extractor, named ICSpk, with higherinterpretability and fewer parameters. Specifically, at first, toquantify the speaker-related frequency response of waveform,we modify the original short-term Fourier transform filters intoa family of complex exponential filters, named interpretablecomplex (IC) filters. Each IC filter is confined by a complexexponential filter parameterized by frequency. Then, a deepcomplex-valued speaker embedding extractor is designed to operateon the complex-valued output of IC filters. The proposedICSpk is evaluated onVoxCeleb andCNCeleb databases. Experimentalresults demonstrate the IC filters-based system exhibitsa significant improvement over the complex spectrogram basedsystems. Furthermore, the proposed ICSpk outperforms existingraw waveform based systems by a large margin.