Chinese speech emotion recognition based on bidirectional long short-term memory network

Hu Du; Kaoru Hirota; Yaping Dai; Donggyun Kim; Junjie Ma

Chinese speech emotion recognition based on bidirectional long short-term memory network

Hu Du, Kaoru Hirota^*, Yaping Dai, Donggyun Kim, Junjie Ma

^*Corresponding author for this work

School of Automation

Beijing Institute of Technology

Research output: Contribution to conference › Paper › peer-review

Abstract

A Bidirectional Long Short-Term Memory(BLSTM) network is applied to improve the accuracy of Chinese speech emotion recognition of six basic human emotions (angry, fear, happy, neutral, sad, and surprise). The features of emotions can be learned and saved by BLSTM network whose special architecture called memory blocks is used to remember information from a long sentence, and BLSTM network provides information both from history and future of the current frame for the importance of the context of sentences. Results of experiments on the CASIA Chinese emotion corpus show that the average recognition accuracy reaches 73.83%, and has a 9.83% increase compared with the method based on information cell, 7.83% increase compared with Mel Frequency Cepstrum Coefficient and Principal Component Analysis, and 24.83% increase compared with Random Deep Belief Networks.

Original language	English
Publication status	Published - 2017
Event	5th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2017 - Beijing, China Duration: 2 Nov 2017 → 5 Nov 2017

Conference

Conference	5th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2017
Country/Territory	China
City	Beijing
Period	2/11/17 → 5/11/17

Keywords

Bidirectional long short-term memory
CASIA Chinese corpus
Chinese speech emotion recognition

Cite this

@conference{8579074ba49449b29d54c48f5a056438,

title = "Chinese speech emotion recognition based on bidirectional long short-term memory network",

abstract = "A Bidirectional Long Short-Term Memory(BLSTM) network is applied to improve the accuracy of Chinese speech emotion recognition of six basic human emotions (angry, fear, happy, neutral, sad, and surprise). The features of emotions can be learned and saved by BLSTM network whose special architecture called memory blocks is used to remember information from a long sentence, and BLSTM network provides information both from history and future of the current frame for the importance of the context of sentences. Results of experiments on the CASIA Chinese emotion corpus show that the average recognition accuracy reaches 73.83%, and has a 9.83% increase compared with the method based on information cell, 7.83% increase compared with Mel Frequency Cepstrum Coefficient and Principal Component Analysis, and 24.83% increase compared with Random Deep Belief Networks.",

keywords = "Bidirectional long short-term memory, CASIA Chinese corpus, Chinese speech emotion recognition",

author = "Hu Du and Kaoru Hirota and Yaping Dai and Donggyun Kim and Junjie Ma",

year = "2017",

language = "English",

note = "5th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2017 ; Conference date: 02-11-2017 Through 05-11-2017",

}

TY - CONF

T1 - Chinese speech emotion recognition based on bidirectional long short-term memory network

AU - Du, Hu

AU - Hirota, Kaoru

AU - Dai, Yaping

AU - Kim, Donggyun

AU - Ma, Junjie

PY - 2017

Y1 - 2017

N2 - A Bidirectional Long Short-Term Memory(BLSTM) network is applied to improve the accuracy of Chinese speech emotion recognition of six basic human emotions (angry, fear, happy, neutral, sad, and surprise). The features of emotions can be learned and saved by BLSTM network whose special architecture called memory blocks is used to remember information from a long sentence, and BLSTM network provides information both from history and future of the current frame for the importance of the context of sentences. Results of experiments on the CASIA Chinese emotion corpus show that the average recognition accuracy reaches 73.83%, and has a 9.83% increase compared with the method based on information cell, 7.83% increase compared with Mel Frequency Cepstrum Coefficient and Principal Component Analysis, and 24.83% increase compared with Random Deep Belief Networks.

AB - A Bidirectional Long Short-Term Memory(BLSTM) network is applied to improve the accuracy of Chinese speech emotion recognition of six basic human emotions (angry, fear, happy, neutral, sad, and surprise). The features of emotions can be learned and saved by BLSTM network whose special architecture called memory blocks is used to remember information from a long sentence, and BLSTM network provides information both from history and future of the current frame for the importance of the context of sentences. Results of experiments on the CASIA Chinese emotion corpus show that the average recognition accuracy reaches 73.83%, and has a 9.83% increase compared with the method based on information cell, 7.83% increase compared with Mel Frequency Cepstrum Coefficient and Principal Component Analysis, and 24.83% increase compared with Random Deep Belief Networks.

KW - Bidirectional long short-term memory

KW - CASIA Chinese corpus

KW - Chinese speech emotion recognition

UR - http://www.scopus.com/inward/record.url?scp=85037522181&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85037522181

T2 - 5th International Workshop on Advanced Computational Intelligence and Intelligent Informatics, IWACIII 2017

Y2 - 2 November 2017 through 5 November 2017

ER -

Chinese speech emotion recognition based on bidirectional long short-term memory network

Abstract

Conference

Keywords

Other files and links

Fingerprint

Cite this