Significance of Accurate Vowel Region Detection for Speech based Emotion Recognition

Speech emotion/affective state recognition alludes to the detection of the human emotive condition from a spoken utterance. Speech emotion classification was studied using the emotion-specific information obtained from various regions of speech signals like utterances, words, syllables, phones, vowel, consonant, unvoiced, and voiced sounds. This paper is aimed to investigate the significance of vowel region detection accuracy for emotion recognition in noisy conditions. Previously, vowel region detection techniques based on vowel onset point and distance criteria were explored for emotion recognition. We study the performance of the nonlinear energy-based accurate vowel-region extraction system integrated into the speech-emotion recognition system. The experiments were performed using features based on single frequency filtering cepstral coefficients combined with i-vector with Probabilistic Linear Discriminant Analysis (PLDA) based classifier, on EMODB and FAU-ABIO databases. Results of the study indicate that an emotion recognition system using speech-features obtained from accurate vowel regions has shown better performance in degraded conditions.