A Comparative Study of the Performance of HMM, DNN, and RNN based Speech Synthesis Systems Trained on Very Large Speaker-Dependent Corpora

Xin Wang; Shinji Takaki; Junichi Yamagishi

doi:10.21437/SSW.2016-20

A Comparative Study of the Performance of HMM, DNN, and RNN based Speech Synthesis Systems Trained on Very Large Speaker-Dependent Corpora

Xin Wang, Shinji Takaki, Junichi Yamagishi

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

Abstract / Description of output

This study investigates the impact of the amount of training data on the performance of parametric speech synthesis systems. A Japanese corpus with 100 hours’ audio recordings of a male voice and another corpus with 50 hours’ recordings of a female voice were utilized to train systems based on hidden Markov model (HMM), feed-forward neural network and recurrent neural network (RNN). The results show that the improvement on the accuracy of the predicted spectral features gradually diminishes as the amount of training data increases. However, different from the “diminishing returns” in the spectral stream, the accuracy of the predicted F0 trajectory by the HMM and RNN systems tends to consistently benefit from the increasing amount of training data.

Original language	English
Title of host publication	Proceedings of 9th ISCA Speech Synthesis Workshop
Pages	125-128
Number of pages	4
DOIs	https://doi.org/10.21437/SSW.2016-20
Publication status	Published - 15 Sept 2016
Event	9th ISCA Speech Synthesis Workshop - Sunnyvale, United States Duration: 13 Sept 2016 → 15 Sept 2016 http://ssw9.talp.cat/

Conference

Conference	9th ISCA Speech Synthesis Workshop
Abbreviated title	ISCA 2016
Country/Territory	United States
City	Sunnyvale
Period	13/09/16 → 15/09/16
Internet address	http://ssw9.talp.cat/

Access to Document

10.21437/SSW.2016-20

SSW9_Xin2Accepted author manuscript, 155 KB

https://www.isca-speech.org/archive/SSW_2016/abstracts/ssw9_OS2-4_Wang.html

Deep architectures for statistical speech synthesis
Yamagishi, J.
UK industry, commerce and public corporations
4/09/12 → 3/03/16
Project: Research
Natural Speech Technology
Renals, S. & King, S.
EPSRC
1/05/11 → 31/07/16
Project: Research

Cite this

@inproceedings{4f5080ce294a404eae0d67b2c0372496,

title = "A Comparative Study of the Performance of HMM, DNN, and RNN based Speech Synthesis Systems Trained on Very Large Speaker-Dependent Corpora",

abstract = "This study investigates the impact of the amount of training data on the performance of parametric speech synthesis systems. A Japanese corpus with 100 hours{\textquoteright} audio recordings of a male voice and another corpus with 50 hours{\textquoteright} recordings of a female voice were utilized to train systems based on hidden Markov model (HMM), feed-forward neural network and recurrent neural network (RNN). The results show that the improvement on the accuracy of the predicted spectral features gradually diminishes as the amount of training data increases. However, different from the “diminishing returns” in the spectral stream, the accuracy of the predicted F0 trajectory by the HMM and RNN systems tends to consistently benefit from the increasing amount of training data.",

author = "Xin Wang and Shinji Takaki and Junichi Yamagishi",

year = "2016",

month = sep,

day = "15",

doi = "10.21437/SSW.2016-20",

language = "English",

pages = "125--128",

booktitle = "Proceedings of 9th ISCA Speech Synthesis Workshop",

note = "9th ISCA Speech Synthesis Workshop , ISCA 2016 ; Conference date: 13-09-2016 Through 15-09-2016",

url = "http://ssw9.talp.cat/",

}

TY - GEN

T1 - A Comparative Study of the Performance of HMM, DNN, and RNN based Speech Synthesis Systems Trained on Very Large Speaker-Dependent Corpora

AU - Wang, Xin

AU - Takaki, Shinji

AU - Yamagishi, Junichi

PY - 2016/9/15

Y1 - 2016/9/15

N2 - This study investigates the impact of the amount of training data on the performance of parametric speech synthesis systems. A Japanese corpus with 100 hours’ audio recordings of a male voice and another corpus with 50 hours’ recordings of a female voice were utilized to train systems based on hidden Markov model (HMM), feed-forward neural network and recurrent neural network (RNN). The results show that the improvement on the accuracy of the predicted spectral features gradually diminishes as the amount of training data increases. However, different from the “diminishing returns” in the spectral stream, the accuracy of the predicted F0 trajectory by the HMM and RNN systems tends to consistently benefit from the increasing amount of training data.

AB - This study investigates the impact of the amount of training data on the performance of parametric speech synthesis systems. A Japanese corpus with 100 hours’ audio recordings of a male voice and another corpus with 50 hours’ recordings of a female voice were utilized to train systems based on hidden Markov model (HMM), feed-forward neural network and recurrent neural network (RNN). The results show that the improvement on the accuracy of the predicted spectral features gradually diminishes as the amount of training data increases. However, different from the “diminishing returns” in the spectral stream, the accuracy of the predicted F0 trajectory by the HMM and RNN systems tends to consistently benefit from the increasing amount of training data.

U2 - 10.21437/SSW.2016-20

DO - 10.21437/SSW.2016-20

M3 - Conference contribution

SP - 125

EP - 128

BT - Proceedings of 9th ISCA Speech Synthesis Workshop

T2 - 9th ISCA Speech Synthesis Workshop

Y2 - 13 September 2016 through 15 September 2016

ER -

A Comparative Study of the Performance of HMM, DNN, and RNN based Speech Synthesis Systems Trained on Very Large Speaker-Dependent Corpora

Abstract / Description of output

Conference

Access to Document

Fingerprint

Projects

Deep architectures for statistical speech synthesis

Natural Speech Technology

Cite this