ExcitNet vocoder: A neural excitation model for parametric speech synthesis systems

Authors: Eunwoo Song, Kyungguen Byun, Hong-Goo Kang
Date: 06 Jun 2019 (last updated 30 Oct 2020)

• Abstract
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.

• Published in: EUSIPCO 2019
• Preprinted version: 1811.04769

Table of Contents

• Speech analysis and synthesis
• Vocoding model: ITFTE LPC vocoder [1], WaveNet vocoder [2], ExcitNet vocoder
• Text-to-speech
• Acoustic model: Parametric model [1]
• Vocoding model: ITFTE LPC vocoder, WaveNet vocoder, ExcitNet vocoder
* Demo for the Tacotron acoustic model with ExcitNet vocoder is [here]

Speech analysis and synthesis

Korean female speaker


RAW
Speaker-dependent model
-	WaveNet Vocoder	ExcitNet Vocoder
Speaker-independent model
ITFTE Vocoder	WaveNet Vocoder	ExcitNet Vocoder


Korean male speaker


RAW
Speaker-dependent model
-	WaveNet Vocoder	ExcitNet Vocoder
Speaker-independent model
ITFTE Vocoder	WaveNet Vocoder	ExcitNet Vocoder

Text-to-speech

Korean female speaker


RAW
Speaker-dependent model
-	WaveNet Vocoder	ExcitNet Vocoder
Speaker-independent model
ITFTE Vocoder	WaveNet Vocoder	ExcitNet Vocoder


Korean male speaker


RAW
Speaker-dependent model
-	WaveNet Vocoder	ExcitNet Vocoder
Speaker-independent model
ITFTE Vocoder	WaveNet Vocoder	ExcitNet Vocoder

References

[1] E. Song, F. K. Soong, and H.-G. Kang, “Effective spectral and excitation modeling techniques for LSTM-RNN-based speech synthesis systems,” IEEE/ACM Trans. Audio, Speech, and Lang. Process., vol. 25, no. 11, pp. 2152–2161, 2017.
[2] K. Tachibana, T. Toda, Y. Shiga, and H. Kawai, “An investigation of noise shaping with perceptual weighting for WaveNet-based speech generation,” in Proc. ICASSP, 2018, pp. 5664–5668.

Acknowledgements

Work performed with nVoice team, Clova Voice, Naver Corp.

Citation

@inproceedings{song2019excitnet,
    title={Excit{N}et vocoder: {A} neural excitation model for parametric speech synthesis systems},
    author={Song, Eunwoo and Byun, Kyungguen and Kang, Hong-Goo},
    booktitle={Proc. EUSIPCO},
    pages={1--5},
    year={2019}
  }