NeuralDPS: Neural Deterministic Plus Stochastic Model with Multiband Excitation for Noise-Controllable Waveform Generation

The traditional vocoders have the advantages of high synthesis efficiency, strong interpretability, and speech editability, while neural advantage quality. To combine two vocoders, inspired by deterministic plus stochastic model, this paper proposes a novel vocoder named NeuralDPS which can retain quality acquire efficiency noise controllability. Firstly, framework contains four modules: source module, V/UV decision module filter module. input required is just spectral parameter, avoids error caused estimating additional parameters, such as F0. Secondly, to solve problem that different frequency bands may proportions components components, multiband excitation strategy used generate more accurate signal reduce filter's burden. Thirdly, method control proposed. In way, signal-to-noise ratio (SNR) be adjusted easily. Objective subjective experimental results show our proposed obtain similar performance with WaveNet it generates waveforms at least 280 times faster than vocoder. It also 28% WaveGAN's on single CPU core. We verified through experiments effectively in predicted adjust SNR speech. Examples generated found https://hairuo55.github.io/NeuralDPS.