SCP-GAN SELF-CORRECTING DISCRIMINATOR OPTIMIZATION FOR TRAINING CONSISTENCY PRESERVING METRIC GAN ON SPEECH ENHANCEMENT TASKS Vasily Zadorozhnyy1 Qiang Ye1 and Kazuhito Koishida2

2025-05-03 0 0 2.32MB 5 页 10玖币

侵权投诉

SCP-GAN: SELF-CORRECTING DISCRIMINATOR OPTIMIZATION FOR TRAINING

CONSISTENCY PRESERVING METRIC GAN ON SPEECH ENHANCEMENT TASKS

Vasily Zadorozhnyy∗,1, Qiang Ye1, and Kazuhito Koishida2

1Department of Mathematics, University of Kentucky, Lexington, USA

2Applied Sciences Group, Microsoft Corporation, Redmond, USA

1{vasily.zadorozhnyy, qye3}@uky.edu 2kazukoi@microsoft.com

ABSTRACT

In recent years, Generative Adversarial Networks (GANs)

have produced signiﬁcantly improved results in speech en-

hancement (SE) tasks. They are difﬁcult to train, however.

In this work, we introduce several improvements to the GAN

training schemes, which can be applied to most GAN-based

SE models. We propose using consistency loss functions,

which target the inconsistency in time and time-frequency

domains caused by Fourier and Inverse Fourier Transforms.

We also present self-correcting optimization for training a

GAN discriminator on SE tasks, which helps avoid “harm-

ful” training directions for parts of the discriminator loss

function. We have tested our proposed methods on several

state-of-the-art GAN-based SE models and obtained consis-

tent improvements, including new state-of-the-art results for

the Voice Bank+DEMAND dataset.

Index Terms—Speech Enhancement, Generative Adver-

sarial Network, MetricGAN, Self-Correcting Optimization,

STFT Consistency, Voice Bank+DEMAND

1. INTRODUCTION

Speech Enhancement (SE) is a process of making deteriorated

speech signals more understandable and perceptually pleas-

ing. The SE has been widely used for various applications, in-

cluding mobile communication, speech recognition systems,

hearing aids, etc. SE as an area of research interest has been

around for several decades. Traditional SE techniques [1,2]

often use a heuristic or straightforward signal processing al-

gorithm to estimate a gain function, which is then applied to

the noisy input to produce improved speech. Recent devel-

opments in deep learning have inspired many Deep Neural

Network (DNN)-based SE techniques [3,4,5,6,7] that out-

perform conventional signal processing-based methods. One

particular DNN-based architecture, Generative Adversarial

Net (GAN), has garnered much interest in the SE commu-

nity for the past few years [5,6,8,9]. In the applications

of SE, GAN architecture is primarily employed to generate

∗Work performed while Vasily Zadorozhnyy was an intern at Microsoft.

enhanced speech. One of the earliest works where GAN mod-

els were implemented on the SE domain is the SEGAN [5]

model. It utilizes an adversarial framework to map the noisy

waveform to a corresponding enhanced speech. Later, Met-

ricGAN [6] introduced a metric score optimization scheme,

where an evaluated metric was introduced into adversarial

loss functions, replacing a traditional binary-classiﬁer [5] and

creating a new branch for SE GAN-based research. There

have been several improvements to the MetricGAN model,

e.g. MetricGAN+ [8], iMetricGAN [10], CMGAN [9], etc.

More recently, with a rise of Transformers [11] and Conform-

ers [12], models such as DB-AIAT [13], DPT-FSNet [14],

SE-Conformer [15], CMGAN [9], etc. show signiﬁcant im-

provements on SE tasks.

Despite much work, training of GAN-based models are

prone to problems such as non-convergence, overﬁtting, and

gradient instabilities. One common issue in GAN’s discrimi-

nator training is potentially “harmful” gradient direction [16]

where parts of the model might train opposite to the desired

direction. To overcome this problem, we propose a new

method called Self-Correcting (SC) Discriminator Optimiza-

tion. At the same time, the SE DNN-based models are subject

to problems caused by the signal-processing tools, e.g., an in-

consistency in the Short-Time Fourier Transform (STFT) and

its inverse (iSTFT) [7,17]. Inspired by [18], we adapt and in-

troduce the consistency loss function as a part of Consistency

Preserving (CP) Net into the GAN framework, where loss

and architecture take into account the iSTFT effects. From

our experiments, the combination of SC and CP methods

improves the SE GAN-based models even further than either

single method; we call such a combination SCP-GAN.

The remainder of this paper is laid out as follows. In sec-

tion 2, we list earlier works pertinent to our work. In section

3, we introduce our improvements to current GAN-based SE

models. We present and compare the SCP-GAN results on

Voice Bank+DEMAND dataset [19] to the current state-of-

the-art (SOTA) models in section 4. Then, in section 5, we

provide an extensive ablation study to show the advantages of

the proposed methods. Finally, in section 6, we highlight the

methods’ contributions to the ﬁeld.

arXiv:2210.14474v1 [cs.SD] 26 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

SCP-GAN:SELF-CORRECTINGDISCRIMINATOROPTIMIZATIONFORTRAININGCONSISTENCYPRESERVINGMETRICGANONSPEECHENHANCEMENTTASKSVasilyZadorozhnyy;1,QiangYe1,andKazuhitoKoishida21DepartmentofMathematics,UniversityofKentucky,Lexington,USA2AppliedSciencesGroup,MicrosoftCorporation,Redmond,USA1fvasily.zadorozhnyy,qye...

展开>> 收起<<

SCP-GAN SELF-CORRECTING DISCRIMINATOR OPTIMIZATION FOR TRAINING CONSISTENCY PRESERVING METRIC GAN ON SPEECH ENHANCEMENT TASKS Vasily Zadorozhnyy1 Qiang Ye1 and Kazuhito Koishida2.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

SCP-GAN SELF-CORRECTING DISCRIMINATOR OPTIMIZATION FOR TRAINING CONSISTENCY PRESERVING METRIC GAN ON SPEECH ENHANCEMENT TASKS Vasily Zadorozhnyy1 Qiang Ye1 and Kazuhito Koishida2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: