Effect of different splitting criteria on the performance of speech emotion recognition Bagus Tris Atmaja

2025-05-03 0 0 215.53KB 5 页 10玖币

侵权投诉

Effect of different splitting criteria on the

performance of speech emotion recognition

Bagus Tris Atmaja*

National Institute of Advanced Industrial Science and Technology, Japan

b-atmaja@aist.go.jp

Akira Sasou

National Institute of Advanced Industrial Science and Technology, Japan

a-sasou@aist.go.jp

Abstract—Traditional speech emotion recognition (SER) eval-

uations have been performed merely on a speaker-independent

condition; some of them even did not evaluate their result on

this condition. This paper highlights the importance of splitting

training and test data for SER by script, known as sentence-

open or text-independent criteria. The results show that em-

ploying sentence-open criteria degraded the performance of

SER. This ﬁnding implies the difﬁculties of recognizing emotion

from speech in different linguistic information embedded in

acoustic information. Surprisingly, text-independent criteria

consistently performed worse than speaker+text-independent

criteria. The full order of difﬁculties for splitting criteria on

SER performances from the most difﬁcult to the easiest is text-

independent, speaker+text-independent, speaker-independent,

and speaker+text-dependent. The gap between speaker+text-

independent and text-independent was smaller than other

criteria, strengthening the difﬁculties of recognizing emotion

from speech in different sentences.

Index Terms—Speech emotion recognition, data partition,

text-independent, speaker-independent, splitting criteria

I. INTRODUCTION

Speech emotion recognition (SER) is one topic of interest

in automatic speech recognition and understanding. In con-

trast to automatic speech recognition (ASR) which attempts

to obtain linguistic information from speech, SER attempts

to obtain non-linguistic information from speech. In more

concrete, SER aims to infer the affective state of the speaker

from solely speech data.

SER can be designed to recognize discrete emotion, con-

tinuous emotion, or both emotion models. Recent research

suggested that emotion is ordinal by nature [1], which is

closer to a categorical than the continuous model. In the

categorical model, several emotion categories exist, from the

simplest two categories with positive and negative emotions

to 27 categories [2]. The choice of emotion model in SER

depends on the availability of the labels in the dataset.

Data-driven methods, in which most SER systems employ

this kind of approach, rely on the conﬁguration or selection

of the data to build the model. In SER, it is common to

split the data by evaluating different speakers for training and

test partitions. This approach, known as speaker-independent

criteria, is a gold standard to build SER model that minimizes

speaker variability in the training phase.

*Corresponding author, on leave from Department of Engineering

Physics, Institut Teknologi Sepuluh Nopember, Surabaya, Indonesia (email:

bagus@ep.its.ac.id).

Recent research in SER, particularly by fusing acoustic

and linguistic information, has found that different splitting

criteria in splitting data for training will result in different

performances [3]. In the fusion of acoustic and linguis-

tic information, it is sound that the model needs to be

trained in different scripts, known as sentence-open or text-

independent. This strategy was intended to avoid the effect of

having the same linguistic information for predicting emotion

under the same sentences for both training and test partitions.

Since linguistic information is extracted via text or script, this

splitting condition is necessary to evaluate such discrepancies

of using different features or types of information. Using

merely acoustic features for SER, one may argue that this

evaluation is unnecessary since no linguistic features are

involved in building the SER model.

Fujisaki in 2002 proposed a scheme in which various types

of information are manifested in the segmental and supraseg-

mental features of speech [4]. One of the types of information

includes emotion. Referencing this argument that emotional

information is manifested directly in speech without a need to

convert speech into text, there is a possibility that different

sentences will yield different SER performances under the

same acoustic-only system. The current research on SER

showed no evaluation of the differences of splitting crite-

ria, particularly comparing data with and without linguistic

information.

The contribution of this paper is an evaluation of the

effect of splitting criteria into the training data on SER

performance. As argued previously, linguistic information

is embedded in acoustic features; hence, evaluating text-

independent criteria, i.e., different sentences for training

and test, is necessary to observe such effects. We eval-

uated four splitting criteria: speaker-dependent (including

text-dependent data), speaker-independent, text-independent,

and speaker+text-independent criteria, and traced their SER

performances. We experimented with these criteria in three

different experiments. The results of three different experi-

ments show a consistent pattern of difﬁculties for four differ-

ent splitting criteria. Text-independent criteria obtained the

worst result followed by speaker+text-independent, speaker-

independent, and speaker-dependent criteria.

II. RELATED WORK

Research on the evaluation of the effect of data selection

on speech processing is not new. Different data for training

arXiv:2210.14501v1 [eess.AS] 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：

EffectofdifferentsplittingcriteriaontheperformanceofspeechemotionrecognitionBagusTrisAtmaja*NationalInstituteofAdvancedIndustrialScienceandTechnology,Japanb-atmaja@aist.go.jpAkiraSasouNationalInstituteofAdvancedIndustrialScienceandTechnology,Japana-sasou@aist.go.jpAbstractTraditionalspeechemotionre...

展开>> 收起<<

Effect of different splitting criteria on the performance of speech emotion recognition Bagus Tris Atmaja.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！

Effect of different splitting criteria on the performance of speech emotion recognition Bagus Tris Atmaja

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: