Classifying emotional states using pitch and formants in vowel regions

In the field of Human Computer Interaction (HCI), human emotion recognition from speech signal is evolving as a recent research area. Speech is the most common way for communication among human beings. Speech consists of sentences, which can be further segregated into words. Words consist of phonemes which are considered to be the primary voice construction elements. This paper presents a classification of four basic emotional states, namely anger, happy, sad, and neutral by extracting acoustic features from the speech signal. Production features mainly F0, i.e., pitch and formants F1, F2, and F3 are derived from the speech signal using only the vowel parts of English language i.e., /a/, /e/, /i/, /o/, and /u/, without requiring to process the speech signal of entire utterances or sentences. Using the pitch and formants feature vectors, the emotion classification has been carried out using a Support Vector Machine (SVM) classifier. In this preliminary investigation, the vowel regions have been separated manually, so as to assess their efficacy in classifying the emotions. The approach has been validated using an emotional speech dataset in English language, collected especially for this study. The performance evaluation results obtained are encouraging. This approach can be further refined for wider applications.