Towards end-to-end synthetic speech detection

Guang Hua; Andrew Beng Jin Teoh; Haijian Zhang

doi:10.1109/LSP.2021.3089437

Towards end-to-end synthetic speech detection

Guang Hua, Andrew Beng Jin Teoh, Haijian Zhang

Department of Electrical and Electronic Engineering

Research output: Contribution to journal › Article › peer-review

19 Citations (Scopus)

Abstract

The constant Q transform (CQT) has been shown to be one of the most effective speech signal pre-transforms to facilitate synthetic speech detection, followed by either hand-crafted (subband) constant Q cepstral coefficient (CQCC) feature extraction and a back-end binary classifier, or a deep neural network (DNN) directly for further feature extraction and classification. Despite the rich literature on such a pipeline, we show in this paper that the pre-transform and hand-crafted features could simply be replaced by end-to-end DNNs. Specifically, we experimentally verify that by only using standard components, a light-weight neural network could outperform the state-of-the-art methods for the ASVspoof2019 challenge. The proposed model is termed Time-domain Synthetic Speech Detection Net (TSSDNet), having ResNet- or Inception-style structures. We further demonstrate that the proposed models also have attractive generalization capability. Trained on ASVspoof2019, they could achieve promising detection performance when tested on disjoint ASVspoof2015, significantly better than the existing cross-dataset results. This paper reveals the great potential of end-to-end DNNs for synthetic speech detection, without hand-crafted features.

Original language	English
Article number	9456037
Pages (from-to)	1265-1269
Number of pages	5
Journal	IEEE Signal Processing Letters
Volume	28
DOIs	https://doi.org/10.1109/LSP.2021.3089437
Publication status	Published - 2021

Bibliographical note

Funding Information:
Manuscript received April 12, 2021; revised June 3, 2021; accepted June 10, 2021. Date of publication June 15, 2021; date of current version June 28, 2021. This work was supported in part by the 2020–2021 International Scholar Exchange Fellowship (ISEF) Program at the Chey Institute for Advanced Studies, South Korea, and in part by the National Natural Science Foundation of China under Grant 61802284. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Yu Tsao. (Corresponding author: Andrew Beng Jin Teoh.) Guang Hua and Haijian Zhang are with the School of Electronic Information, Wuhan University, Wuhan 430072, China (e-mail: ghua@whu.edu.cn; haijian.zhang@whu.edu.cn).

Publisher Copyright:
© 2021 IEEE.

All Science Journal Classification (ASJC) codes

Signal Processing
Applied Mathematics
Electrical and Electronic Engineering

Access to Document

10.1109/LSP.2021.3089437

Cite this

@article{f8d34de3dac34319bcc769df4c5a6a88,

title = "Towards end-to-end synthetic speech detection",

abstract = "The constant Q transform (CQT) has been shown to be one of the most effective speech signal pre-transforms to facilitate synthetic speech detection, followed by either hand-crafted (subband) constant Q cepstral coefficient (CQCC) feature extraction and a back-end binary classifier, or a deep neural network (DNN) directly for further feature extraction and classification. Despite the rich literature on such a pipeline, we show in this paper that the pre-transform and hand-crafted features could simply be replaced by end-to-end DNNs. Specifically, we experimentally verify that by only using standard components, a light-weight neural network could outperform the state-of-the-art methods for the ASVspoof2019 challenge. The proposed model is termed Time-domain Synthetic Speech Detection Net (TSSDNet), having ResNet- or Inception-style structures. We further demonstrate that the proposed models also have attractive generalization capability. Trained on ASVspoof2019, they could achieve promising detection performance when tested on disjoint ASVspoof2015, significantly better than the existing cross-dataset results. This paper reveals the great potential of end-to-end DNNs for synthetic speech detection, without hand-crafted features.",

author = "Guang Hua and Teoh, {Andrew Beng Jin} and Haijian Zhang",

note = "Funding Information: Manuscript received April 12, 2021; revised June 3, 2021; accepted June 10, 2021. Date of publication June 15, 2021; date of current version June 28, 2021. This work was supported in part by the 2020–2021 International Scholar Exchange Fellowship (ISEF) Program at the Chey Institute for Advanced Studies, South Korea, and in part by the National Natural Science Foundation of China under Grant 61802284. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Yu Tsao. (Corresponding author: Andrew Beng Jin Teoh.) Guang Hua and Haijian Zhang are with the School of Electronic Information, Wuhan University, Wuhan 430072, China (e-mail: ghua@whu.edu.cn; haijian.zhang@whu.edu.cn). Publisher Copyright: {\textcopyright} 2021 IEEE.",

year = "2021",

doi = "10.1109/LSP.2021.3089437",

language = "English",

volume = "28",

pages = "1265--1269",

journal = "IEEE Signal Processing Letters",

issn = "1070-9908",

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

}

TY - JOUR

T1 - Towards end-to-end synthetic speech detection

AU - Hua, Guang

AU - Teoh, Andrew Beng Jin

AU - Zhang, Haijian

N1 - Funding Information: Manuscript received April 12, 2021; revised June 3, 2021; accepted June 10, 2021. Date of publication June 15, 2021; date of current version June 28, 2021. This work was supported in part by the 2020–2021 International Scholar Exchange Fellowship (ISEF) Program at the Chey Institute for Advanced Studies, South Korea, and in part by the National Natural Science Foundation of China under Grant 61802284. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Yu Tsao. (Corresponding author: Andrew Beng Jin Teoh.) Guang Hua and Haijian Zhang are with the School of Electronic Information, Wuhan University, Wuhan 430072, China (e-mail: ghua@whu.edu.cn; haijian.zhang@whu.edu.cn). Publisher Copyright: © 2021 IEEE.

PY - 2021

Y1 - 2021

N2 - The constant Q transform (CQT) has been shown to be one of the most effective speech signal pre-transforms to facilitate synthetic speech detection, followed by either hand-crafted (subband) constant Q cepstral coefficient (CQCC) feature extraction and a back-end binary classifier, or a deep neural network (DNN) directly for further feature extraction and classification. Despite the rich literature on such a pipeline, we show in this paper that the pre-transform and hand-crafted features could simply be replaced by end-to-end DNNs. Specifically, we experimentally verify that by only using standard components, a light-weight neural network could outperform the state-of-the-art methods for the ASVspoof2019 challenge. The proposed model is termed Time-domain Synthetic Speech Detection Net (TSSDNet), having ResNet- or Inception-style structures. We further demonstrate that the proposed models also have attractive generalization capability. Trained on ASVspoof2019, they could achieve promising detection performance when tested on disjoint ASVspoof2015, significantly better than the existing cross-dataset results. This paper reveals the great potential of end-to-end DNNs for synthetic speech detection, without hand-crafted features.

AB - The constant Q transform (CQT) has been shown to be one of the most effective speech signal pre-transforms to facilitate synthetic speech detection, followed by either hand-crafted (subband) constant Q cepstral coefficient (CQCC) feature extraction and a back-end binary classifier, or a deep neural network (DNN) directly for further feature extraction and classification. Despite the rich literature on such a pipeline, we show in this paper that the pre-transform and hand-crafted features could simply be replaced by end-to-end DNNs. Specifically, we experimentally verify that by only using standard components, a light-weight neural network could outperform the state-of-the-art methods for the ASVspoof2019 challenge. The proposed model is termed Time-domain Synthetic Speech Detection Net (TSSDNet), having ResNet- or Inception-style structures. We further demonstrate that the proposed models also have attractive generalization capability. Trained on ASVspoof2019, they could achieve promising detection performance when tested on disjoint ASVspoof2015, significantly better than the existing cross-dataset results. This paper reveals the great potential of end-to-end DNNs for synthetic speech detection, without hand-crafted features.

UR - http://www.scopus.com/inward/record.url?scp=85112110314&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85112110314&partnerID=8YFLogxK

U2 - 10.1109/LSP.2021.3089437

DO - 10.1109/LSP.2021.3089437

M3 - Article

AN - SCOPUS:85112110314

SN - 1070-9908

VL - 28

SP - 1265

EP - 1269

JO - IEEE Signal Processing Letters

JF - IEEE Signal Processing Letters

M1 - 9456037

ER -

Towards end-to-end synthetic speech detection

Abstract

Bibliographical note

All Science Journal Classification (ASJC) codes

Access to Document

Other files and links

Fingerprint

Cite this