ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis

Suhyeon Oh; Hyungseob Lim; Kyungguen Byun; Min Jae Hwang; Eunwoo Song; Hong Goo Kang

ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis

Suhyeon Oh, Hyungseob Lim, Kyungguen Byun, Min Jae Hwang, Eunwoo Song, Hong Goo Kang

Department of Electrical and Electronic Engineering

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

2 Citations (Scopus)

Abstract

In this paper, we propose ExcitGlow, a vocoder that incorporates the source-filter model of voice production theory into a flow-based deep generative model. By targeting the distribution of the excitation signal instead of the speech waveform itself, we significantly reduce the size of the flow-based generative model. To further reduce the number of parameters, we apply a parameter sharing technique in which a single affine coupling layer is used for several flow layers. To avoid quality degradation, we also introduce a closed-loop training framework to optimize the flow model for both the speech and excitation signal generation processes. Specifically, we choose negative log-likelihood (NLL) loss for the excitation signal and multi-resolution spectral distance for the speech signal. As a result, we are able to reduce the model size from 87. 73M to 15. 60M parameters while maintaining the perceptual quality of synthesized speech.

Original language	English
Title of host publication	2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings
Publisher	Institute of Electrical and Electronics Engineers Inc.
Pages	831-836
Number of pages	6
ISBN (Electronic)	9789881476883
Publication status	Published - 2020 Dec 7
Event	2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Virtual, Auckland, New Zealand Duration: 2020 Dec 7 → 2020 Dec 10

Publication series

Name	2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings

Conference

Conference	2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020
Country/Territory	New Zealand
City	Virtual, Auckland
Period	20/12/7 → 20/12/10

Bibliographical note

Funding Information:
VII. ACKNOWLEDGEMENTS The work was supported by Clova Voice, NAVER Corp., Seongnam, Korea.

Publisher Copyright:
© 2020 APSIPA.

All Science Journal Classification (ASJC) codes

Artificial Intelligence
Computer Networks and Communications
Computer Vision and Pattern Recognition
Hardware and Architecture
Signal Processing
Decision Sciences (miscellaneous)
Instrumentation

Cite this

Oh, S., Lim, H., Byun, K., Hwang, M. J., Song, E., & Kang, H. G. (2020). ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis. In 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings (pp. 831-836). Article 9306395 (2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings). Institute of Electrical and Electronics Engineers Inc..

Oh, Suhyeon ; Lim, Hyungseob ; Byun, Kyungguen et al. / ExcitGlow : Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis. 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings. Institute of Electrical and Electronics Engineers Inc., 2020. pp. 831-836 (2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings).

@inproceedings{717e5839024c4799994a7140383d2233,

title = "ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis",

abstract = "In this paper, we propose ExcitGlow, a vocoder that incorporates the source-filter model of voice production theory into a flow-based deep generative model. By targeting the distribution of the excitation signal instead of the speech waveform itself, we significantly reduce the size of the flow-based generative model. To further reduce the number of parameters, we apply a parameter sharing technique in which a single affine coupling layer is used for several flow layers. To avoid quality degradation, we also introduce a closed-loop training framework to optimize the flow model for both the speech and excitation signal generation processes. Specifically, we choose negative log-likelihood (NLL) loss for the excitation signal and multi-resolution spectral distance for the speech signal. As a result, we are able to reduce the model size from 87. 73M to 15. 60M parameters while maintaining the perceptual quality of synthesized speech.",

author = "Suhyeon Oh and Hyungseob Lim and Kyungguen Byun and Hwang, {Min Jae} and Eunwoo Song and Kang, {Hong Goo}",

note = "Funding Information: VII. ACKNOWLEDGEMENTS The work was supported by Clova Voice, NAVER Corp., Seongnam, Korea. Publisher Copyright: {\textcopyright} 2020 APSIPA.; 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 ; Conference date: 07-12-2020 Through 10-12-2020",

year = "2020",

month = dec,

day = "7",

language = "English",

series = "2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

pages = "831--836",

booktitle = "2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings",

address = "United States",

}

Oh, S, Lim, H, Byun, K, Hwang, MJ, Song, E & Kang, HG 2020, ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis. in 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings., 9306395, 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings, Institute of Electrical and Electronics Engineers Inc., pp. 831-836, 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020, Virtual, Auckland, New Zealand, 20/12/7.

ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis. / Oh, Suhyeon; Lim, Hyungseob; Byun, Kyungguen et al.
2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings. Institute of Electrical and Electronics Engineers Inc., 2020. p. 831-836 9306395 (2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - ExcitGlow

T2 - 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020

AU - Oh, Suhyeon

AU - Lim, Hyungseob

AU - Byun, Kyungguen

AU - Hwang, Min Jae

AU - Song, Eunwoo

AU - Kang, Hong Goo

PY - 2020/12/7

Y1 - 2020/12/7

N2 - In this paper, we propose ExcitGlow, a vocoder that incorporates the source-filter model of voice production theory into a flow-based deep generative model. By targeting the distribution of the excitation signal instead of the speech waveform itself, we significantly reduce the size of the flow-based generative model. To further reduce the number of parameters, we apply a parameter sharing technique in which a single affine coupling layer is used for several flow layers. To avoid quality degradation, we also introduce a closed-loop training framework to optimize the flow model for both the speech and excitation signal generation processes. Specifically, we choose negative log-likelihood (NLL) loss for the excitation signal and multi-resolution spectral distance for the speech signal. As a result, we are able to reduce the model size from 87. 73M to 15. 60M parameters while maintaining the perceptual quality of synthesized speech.

AB - In this paper, we propose ExcitGlow, a vocoder that incorporates the source-filter model of voice production theory into a flow-based deep generative model. By targeting the distribution of the excitation signal instead of the speech waveform itself, we significantly reduce the size of the flow-based generative model. To further reduce the number of parameters, we apply a parameter sharing technique in which a single affine coupling layer is used for several flow layers. To avoid quality degradation, we also introduce a closed-loop training framework to optimize the flow model for both the speech and excitation signal generation processes. Specifically, we choose negative log-likelihood (NLL) loss for the excitation signal and multi-resolution spectral distance for the speech signal. As a result, we are able to reduce the model size from 87. 73M to 15. 60M parameters while maintaining the perceptual quality of synthesized speech.

UR - http://www.scopus.com/inward/record.url?scp=85100918189&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85100918189&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:85100918189

T3 - 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings

SP - 831

EP - 836

BT - 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings

PB - Institute of Electrical and Electronics Engineers Inc.

Y2 - 7 December 2020 through 10 December 2020

ER -

Oh S, Lim H, Byun K, Hwang MJ, Song E, Kang HG. ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis. In 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings. Institute of Electrical and Electronics Engineers Inc. 2020. p. 831-836. 9306395. (2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2020 - Proceedings).

ExcitGlow: Improving a WaveGlow-based Neural Vocoder with Linear Prediction Analysis

Abstract

Publication series

Conference

Bibliographical note

All Science Journal Classification (ASJC) codes

Other files and links

Fingerprint

Cite this