Speech Emotion Recognition based on Voiced Emotion Unit

Abstract

Speech emotion recognition (SER) system is becoming a very important tool for human-computer interaction. Previous studies in SER have been focused on utterance as one unit. They assumed that the emotional state is fixed during the utterance, although, the emotional state may change during the time even in one utterance. Therefore, using utterance as one unit is not suitable for this purpose especially for long utterances. The ultimate goal of this study is to find a novel emotion unit that can be used to improve SER accuracy. Therefore, different emotion units defined based on voiced segments are investigated. To find the optimal emotion unit, SER system based on support vector machine (SVM) classifier is used to evaluate each unit. The classification rate is used as a metric for the evaluation. To validate the proposed method, the Berlin database of emotional speech EMO-DB is used. The experimental results revealed that emotion unit that contains four voiced segments gives the highest recognition rate for SER. Moreover, the final emotional state of the whole utterance is determined by majority voting of emotional states of its units. It is found that the performance of the proposed method using voiced related emotion unit outperforms the conventional method using utterance as one unit. General TermsPattern Recognition, Machine learning, speech processing.

Breaking the Code: An Investigation on The Vocal Registers of Malaysia’s National Songstress Puan Sri Saloma

“…The researchers chose not to analyse using scientific vocal analysis like Elbarougy (2019) and Wei Shean Ter and Kwan Yie Wong (2022) did. This is because it is not used as widely in vocal pedagogy as the method proposed in this study as a scientific vocal analysis requires various recording devices such as microphones, speakers, and amplifiers, which could only be found in a recording studio.…”