ExcitNet vocoder: A neural excitation model for parametric speech synthesis systems
This work addresses noisy speech synthesis for parametric speech synthesis systems, offering an incremental improvement over existing WaveNet-based methods.
The paper tackled the problem of noisy outputs in WaveNet-based neural vocoding for speech synthesis by proposing ExcitNet, a neural excitation model that decouples spectral components using an adaptive inverse filter and generates the residual excitation signal with WaveNet, resulting in improved speech quality that outperforms traditional linear prediction vocoders and conventional WaveNet vocoders in experiments.
This paper proposes a WaveNet-based neural excitation model (ExcitNet) for statistical parametric speech synthesis systems. Conventional WaveNet-based neural vocoding systems significantly improve the perceptual quality of synthesized speech by statistically generating a time sequence of speech waveforms through an auto-regressive framework. However, they often suffer from noisy outputs because of the difficulties in capturing the complicated time-varying nature of speech signals. To improve modeling efficiency, the proposed ExcitNet vocoder employs an adaptive inverse filter to decouple spectral components from the speech signal. The residual component, i.e. excitation signal, is then trained and generated within the WaveNet framework. In this way, the quality of the synthesized speech signal can be further improved since the spectral component is well represented by a deep learning framework and, moreover, the residual component is efficiently generated by the WaveNet framework. Experimental results show that the proposed ExcitNet vocoder, trained both speaker-dependently and speaker-independently, outperforms traditional linear prediction vocoders and similarly configured conventional WaveNet vocoders.