Feature Extraction of Assamese Speech Based One Motion Analysis

Abstract

Objectives:The present work aims to investigate the recognition of emotion from Assamese speech. Methods: This work presents a method based on the Gaussian Mixture Model (GMM) classifier and Mel-frequency cepstral coefficients (MFCC) as feature extraction technique for emotion recognition from Assamese speeches. Findings: We have conducted experiments considering different emotions: Angry, Happy, Neutral and Sad. The speech emotion recognition system database is the emotional speech samples collected manually from 20 speakers and some standard samples available on the internet. The speakers are from different districts of Assam and use different dialects of the Assamese language, such as Western (Kamrupi), Central, and Eastern. They fall under the age group of 18-26 years. The field survey consists of recordings done at Dibrugarh University and outside the campus. After the GMM training and testing process, the accuracy we obtained is 51.25%. The experiments confirmed that angry and happy emotions have high energy in the higher frequency region. In contrast, neutral and sad emotions have low energy in the higher frequency region. Novelty: This work will help predict the attitudes and actions of different speakers based on their manner of speaking. In addition, the present work will also help in other aspects of human-machine interaction in our daily life. The Assamese emotional speech database used in the work is self-collected from different dialect groups to understand the variability of emotions in dialectal perspective.