Speech Prosody Extraction for Ibibio Emotions Analysis and Classification

Abstract

This paper investigates basic prosodic features of speech (duration, pitch/F0 and intensity) for the analysis and classification of Ibibio (New Benue Congo, Nigeria) emotions, at the suprasegnental (sentence, word, and syllable) level. We begin by proposing generic hypothesis/baselines, representing a cache of research works documented over the years on emotion effects of neutral speech on western languages, and adopt the circumplex model for the effective representation of emotions. Our methodology uses standard approach to speech processing and exploits machine learning for the classification of seven emotions (anger, fear, joy, normal, pride, sadness, and surprise) obtained from male and female speakers. Analysis of feature-emotion correlates reveal that syllable (duration) units yield the least standard deviation across selected emotions for both genders, compared to other units. Also, there appear to be consistency for word and syllable units – as both genders show same duration correlate patterns. For F0 and intensity features, our findings agree with the literature, as high activation effects tend to produce higher F0 and intensity values, compared to low activation effects, but neutral and low activation effects produce the lowest pitch/F0 and intensity values (for both genders). A classification of the emotions yields interesting results, as classification accuracies and errors remarkably improved in emotion-F0 and emotion-intensity classification for support vector machine (SVM) and decision tree (DT) classifiers, but the highest classification accuracies were produced by the three classifiers at the sentence unit/level for fear emotion, with the k-nearest neighbour classifier (k-NN) leading (DT: 90%, SVM: 90%, k-NN: 92.40%).