EVIDENCE OF VOCAL TRACT ARTICULATION IN SELF-SUPERVISED LEARNING OF SPEECH Cheol Jun Cho1Peter Wu1Abdelrahman Mohamed2Gopala K. Anumanchipalli1

2025-05-06 0 0 937.62KB 5 页 10玖币

侵权投诉

EVIDENCE OF VOCAL TRACT ARTICULATION IN SELF-SUPERVISED LEARNING OF

SPEECH

Cheol Jun Cho1Peter Wu1Abdelrahman Mohamed2Gopala K. Anumanchipalli1

1UC Berkeley, EECS, CA

2Meta AI

ABSTRACT

Recent self-supervised learning (SSL) models have proven to

learn rich representations of speech, which can readily be uti-

lized by diverse downstream tasks. To understand such util-

ities, various analyses have been done for speech SSL mod-

els to reveal which and how information is encoded in the

learned representations. Although the scope of previous anal-

yses is extensive in acoustic, phonetic, and semantic perspec-

tives, the physical grounding by speech production has not yet

received full attention. To bridge this gap, we conduct a com-

prehensive analysis to link speech representations to articula-

tory trajectories measured by electromagnetic articulography

(EMA). Our analysis is based on a linear probing approach

where we measure articulatory score as an average correla-

tion of linear mapping to EMA. We analyze a set of SSL

models selected from the leaderboard of the SUPERB bench-

mark [1] and perform further layer-wise analyses on two most

successful models, Wav2Vec 2.0 [2] and HuBERT [3]. Sur-

prisingly, representations from the recent speech SSL models

are highly correlated with EMA traces (best: r= 0.81), and

only 5 minutes are sufﬁcient to train a linear model with high

performance (r= 0.77). Our ﬁndings suggest that SSL mod-

els learn to align closely with continuous articulations, and

provide a novel insight into speech SSL.

Index Terms—Speech, Self-supervised learning, Elec-

tromagnetic articulography (EMA), Speech representation,

Probing analysis, Acoustic-to-articulatory inversion

1. INTRODUCTION

Self-supervised learning (SSL) has been suggested as a pre-

training method to learn representations without requiring a

huge amount of labeled data. Recently proposed SSL models

for speech provide rich representations which can be readily

utilized for a broad range of spoken language tasks. When

ﬁne-tuned to downstream tasks, the SSL-based models are

able to surpass supervised-only models [4, 1]. Understand-

ing how SSL models work is crucial to explain such success

and to improve speech SSL. Previous studies have revealed

acoustic, phonetic, and semantic representations encoded in

representations of speech SSL, using diverse analytic tools in-

Pre-trained SSL model

representations EMA traces

Linear probe

Upper Lip (UL)

Lower Lip (LL)

Lower Incisor (LI)

Tongue Tip (TT)

Tongue Blade (TB)

Tongue Dorsum(TD)

Fig. 1. General framework of our analysis approach: linear

probing of representations from pre-trained SSL models on

EMA

cluding (non-) linear probing, mutual information, or canon-

ical correlation analysis [5, 3, 6, 7, 8]. Although those anal-

yses have provided insights of information process in speech

SSL models, the models still largely remain black box. Since

the advent of data-driven deep learning strategies for spoken

language engineering, they have grown increasingly discon-

nected from human mechanisms and insights from speech

production.

To bridge this gap, we answer the question of how well

the SSL representations are aligned with principles of speech

production. We conduct a linear probing analysis of SSL rep-

resentations against electromagnetic articulography (EMA)

(Fig. 1). EMA measures real-time continuous displacements

of 6 articulators (Fig. 1), which are timely synced with the

produced speech [9, 10]. As EMA tracks the actual phys-

ical dynamics of vocal tract, the modality is suitable for

investigating physical grounding of speech representation.

Moreover, a large portion of speech features are naturally

subsumed by articulatory trajectories and a full speech can be

reconstructed from EMA [11]. Underlying neurobiological

process of speech production can also be explained by vocal

tract articulation [12]. Therefore, we target EMA traces as

principled and grounded reference for probing representa-

tions of SSL models. We ﬁrstly introduce articulatory score

as an average correlation of linear prediction and EMA. Then,

we conduct a comprehensive analysis by controlling model

space, data size, and layers for probing. Our work bridges

arXiv:2210.11723v3 [eess.AS] 21 Jul 2023

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：

EVIDENCEOFVOCALTRACTARTICULATIONINSELF-SUPERVISEDLEARNINGOFSPEECHCheolJunCho1PeterWu1AbdelrahmanMohamed2GopalaK.Anumanchipalli11UCBerkeley,EECS,CA2MetaAIABSTRACTRecentself-supervisedlearning(SSL)modelshaveproventolearnrichrepresentationsofspeech,whichcanreadilybeuti-lizedbydiversedownstreamtasks.Tou...

展开>> 收起<<

EVIDENCE OF VOCAL TRACT ARTICULATION IN SELF-SUPERVISED LEARNING OF SPEECH Cheol Jun Cho1Peter Wu1Abdelrahman Mohamed2Gopala K. Anumanchipalli1.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！

EVIDENCE OF VOCAL TRACT ARTICULATION IN SELF-SUPERVISED LEARNING OF SPEECH Cheol Jun Cho1Peter Wu1Abdelrahman Mohamed2Gopala K. Anumanchipalli1

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: