An Experimental Investigation on Excitation Representation of WaveNet-Based Neural Vocoders

In this paper, we investigate the performance of using different excitation parameters for WaveNet-based neural vocoders by experiments. The neural vocoders based on WaveNet models have been proposed recently and achieved high quality of reconstructed speech. In these vocoders, spectral features and excitation parameters are used as local conditions of WaveNet models to predict the distribution of waveform samples. In our experiments, F0s, pulse trains, and LPC residual signals are utilized to represent excitation parameters separately or jointly. When using pulse trains and LPC residual signals, additional dilated causal convolution blocks are adopted to derive the condition vectors. Experimental results show that the vocoders conditioned on LPC residual signals can achieve better temporal accuracy of waveform modeling and prediction than using F0s and pulse trains as excitation parameters. These results imply the deficiency of only using F0s as excitation parameters and other parameters that can provide more detailed description of excitations should be explored in future work.