Wav2Spk: A Simple DNN Architecture for Learning Speaker Embeddings from Waveforms

Speaker recognition has seen impressive advances with the advent of deep neural networks (DNNs). However, state-of-the-art speaker recognition systems still rely on human engineering features such as mel-frequency cepstrum coefficients (MFCC). We believe that the handcrafted features limit the potential of the powerful representation of DNNs. Besides, there are also additional steps such as voice activity detection (VAD) and cepstral mean and variance normalization (CMVN) after computing the MFCC. In this paper, we show that MFCC, VAD, and CMVN can be replaced by the tools available in the standard deep learning toolboxes, such as a stacked of stride convolutions, temporal gating, and instance normalization. With these tools, we show that directly learning speaker embeddings from waveforms outperforms an x-vector network that uses MFCC or filter-bank output as features. We achieve an EER of 1.95% on the VoxCeleb1 test set using an end-to-end training scheme, which, to our best knowledge, is the best performance reported using raw waveforms. What's more, the proposed method is complementary with x-vector systems. The fusion of the proposed method with x-vectors trained on filter-bank features produce an EER of 1.55%.