Quasi-Periodic Parallel WaveGAN: A Non-Autoregressive Raw Waveform Generative Model With Pitch-Dependent Dilated Convolution Neural Network

In this paper, we propose a quasi-periodic parallel WaveGAN (QPPWG) waveform generative model, which applies (QP) structure to (PWG) model using pitch-dependent dilated convolution networks (PDCNNs). PWG is small-footprint GAN-based raw whose generation time much faster than real because of its compact and non-autoregressive (non-AR) non-causal mechanisms. Although achieves high-fidelity speech generation, the generic simple network architecture lacks pitch controllability for an unseen auxiliary fundamental frequency ($F_{0}$) feature such as scaled $F_{0}$. To improve modeling capability, apply QP with PDCNNs PWG, introduces information by dynamically changing corresponding $F_{0}$ feature. Both objective subjective experimental results show that QPPWG outperforms when scaled. Moreover, analyses intermediate outputs also better tractability interpretability QPPWG, respectively models spectral excitation-like signals cascaded fixed adaptive blocks structure.