The DKU-Tencent System for the VoxCeleb Speaker Recognition Challenge 2022 Xiaoyi Qin12Na Li2 Yuke Lin1 Yiwei Ding2 Chao Weng2 Dan Su2 Ming Li1

2025-05-06 0 0 322.64KB 4 页 10玖币

侵权投诉

The DKU-Tencent System for the VoxCeleb Speaker Recognition Challenge

2022

Xiaoyi Qin1,2∗Na Li2∗, Yuke Lin1∗, Yiwei Ding2, Chao Weng2, Dan Su2, Ming Li1

1Data Science Research Center, Duke Kunshan University, Kunshan, China

2Tencent AI Lab, Shenzhen, China

ming.li@whu.edu.cn, 2020102110042@whu.edu.cn

Abstract

This paper is the system description of the DKU-Tencent

System for the VoxCeleb Speaker Recognition Challenge 2022

(VoxSRC22). In this challenge, we focus on track1 and track3.

For track1, multiple backbone networks are adopted to extract

frame-level features. Since track1 focus on the cross-age sce-

narios, we adopt the cross-age trials and perform QMF to cal-

ibrate score. The magnitude-based quality measures achieve

a large improvement. For track3, the semi-supervised domain

adaptation task, the pseudo label method is adopted to make do-

main adaptation. Considering the noise labels in clustering, the

ArcFace is replaced by Sub-center ArcFace. The ﬁnal submis-

sion achieves 0.107 mDCF in task1 and 7.135% EER in task3.

Index Terms: speech recognition, cross-age, semi-supervised,

domain-adaptation

1. Task1: Fully supervised speaker

veriﬁcation

1.1. Data Augmentation

We adopt the on-the-ﬂy data augmentation [1] to add additive

background noise or convolutional reverberation noise for the

time-domain waveform. The MUSAN [2] and RIR Noise [3]

datasets are used as noise sources and room impulse response

functions, respectively. The convolution operation is performed

for the reverberation with 40,000 simulated room impulse re-

sponses. We only use RIRs from small and medium rooms.

To further diversify training samples, we apply ampliﬁcation

or playback speed change (pitch remains untouched) to audio

signals. Also, we apply speaker augmentation with speed per-

turbation [4]. We speed up or down each utterance by a factor

of 0.9 or 1.1, yielding shifted pitch utterances that are consid-

ered from new speakers. As a result, the training data includes

3,276,027(1,092,009 ×3) utterances from 17,982(5,994 ×

3) speakers. The probability of Noise/RIR/Tempo/Volume is

0.66.

1.2. Speaker Embedding Model

Generally, the speaker embedding model consists of three parts:

backbone network, encoding layer and loss function. The Ar-

cFace [5], which could increase intra-speaker distances while

ensuring inter-speaker compactness, is used as a loss function.

The backbone and encoding layer are reported as follows.

* Equal contribution. The work was done at Tencent during the

internship for the ﬁrst author

1.2.1. Backbone Network

In this module, we introduce four different speaker ver-

iﬁcation systems, including the SimAM-ResNet34[6], the

ResNet101[7], ResNet152, and Res2Net101[8]. The acoustic

features are 80-dimensional log Mel-ﬁlterbank energies with a

frame length of 25ms and hop size of 10ms. The extracted fea-

tures are mean-normalized before feeding into the deep speaker

network.

1.2.2. Encoding Layer

The encoding layer could transform the variable length feature

maps into a ﬁxed dimension vector, that is, the frame-level fea-

ture is converted to the segment-level feature. In this module,

we adopt the statistic pooling[9], standard deviation pooling

and attentive statistic pooling[10]. The statistic pooling layer

computes the mean and standard deviation of the output feature

maps along with the frequency dimension. The standard devia-

tion pooling layer only calculates the standard deviation.

1.3. Training Strategy and Implement Detail

The speaker embedding model is trained through two stages. In

the ﬁrst stage, the baseline system adopts the SGD optimizer.

We adopt the ReduceLROnPlateau scheduler with 2 epochs of

patience. The init learning rate starts from 0.1, the minimum

learning rate is 1.0e-4, and the decay factor is 0.1. The margin

and scale of ArcFace are set as 0.2 and 32, respectively. We per-

form a linear warmup learning rate schedule at the ﬁrst 2 epochs

to prevent model vibration and speed model training. The input

frame length is ﬁxed at 200 frames. In the second stage, we

following the SpeakIn training protocol[11] with large margin

ﬁne-tune (LMFT)[12]. In the LMFT stage, we also remove the

speaker augmentation and reduce the probability of data aug-

mentation to 0.33. What’s more, the margin is increased from

0.2 to 0.5. According to the speaker embedding model size and

the memory limit of GPU, the frame length is expended from

200 to 400 or 600. The learning rate decays from 1.0e-4 to

1.0e-5.

1.4. Score Calibration and Normalization

The cosine similarity is the back-end scoring method. After

scoring, results from all trials are subject to score normaliza-

tion. We utilize Adaptive Symmetric Score Normalization (AS-

Norm) [13] in our systems. The imposter cohort consists of the

average of the length normalized utterance-based embeddings

of each training speaker, i.e., a speaker-wise cohort with 5994

embeddings.

Quality Measure Functions could calibrate the scores and

improve the system performance. As [14] described, we adopt

the two qualities set: speech duration and magnitude rate. For

arXiv:2210.05092v1 [cs.SD] 11 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：

TheDKU-TencentSystemfortheVoxCelebSpeakerRecognitionChallenge2022XiaoyiQin1;2NaLi2,YukeLin1,YiweiDing2,ChaoWeng2,DanSu2,MingLi11DataScienceResearchCenter,DukeKunshanUniversity,Kunshan,China2TencentAILab,Shenzhen,Chinaming.li@whu.edu.cn,2020102110042@whu.edu.cnAbstractThispaperisthesystemdescripti...

展开>> 收起<<

The DKU-Tencent System for the VoxCeleb Speaker Recognition Challenge 2022 Xiaoyi Qin12Na Li2 Yuke Lin1 Yiwei Ding2 Chao Weng2 Dan Su2 Ming Li1.pdf

共4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！

The DKU-Tencent System for the VoxCeleb Speaker Recognition Challenge 2022 Xiaoyi Qin12Na Li2 Yuke Lin1 Yiwei Ding2 Chao Weng2 Dan Su2 Ming Li1

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: