Utilizing Third-Order Hidden Markov Models for Emotional Talking Condition Recognition

Abstract

It is well known that the performance of emotional talking condition recognition is imperfect. This work aims at enhancing the performance of emotional talking condition recognition based on Third-Order Hidden Markov Models (HMM3s) as classifiers. Our work has been tested on Emotional Prosody Speech and Transcripts (EPST) database. The extracted features of EPST database are Mel-Frequency Cepstral Coefficients (MFCCs). Our results give average talking recognition performance of 71.8%. The results of this work show that HMM3s are superior to First-Order Hidden Markov Models (HMM1s) and Second-Order Hidden Markov Models (HMM2s) by 14.0% and 5.7%, respectively, for emotional talking condition recognition. The average talking recognition performance attained based on HMM3s is very close to that obtained using subjective assessment by human judges.