Upgrading the Performance of Speech Emotion Recognition at the Segmental Level

This paper presents an efficient approach for maximizing the accuracy of automatic speech emotion recognition in English, using minimal inputs, minimal features, lesser algorithmic complexity and reduced processing time. Whereas the findings reported here are based on the exclusive use of vowel formants, most of the related previous works used tens or even hundreds of other features. In spite of using a greater level of signal processing, the recognition accuracy reported earlier was often lesser than that obtained by our approach. This method is based on vowel utterances and the first step comprises statistical pre-processing of the vowel formants. This is followed by the identification of the best formants using the KMeans, K-nearest neighbor and Naive Bayes classifiers. The Artificial neural network that was used for the final classification gave an accuracy of 95.6% on elicited emotional speech. Nearly 1500 speech files from ten female speakers in the neutral and six basic emotions were used to prove the efficiency of the proposed approach. Such a result has not been reported earlier for English and is of significance to researchers, sociologists and others interested in speech.