ENTROPY BASED AUDIO EMOTION ANALYSIS

Abstract

Humans commonly interact with each other using speech. Extracting information from the speech helps in effective interaction between humans and computers. Thus analyzing and recognizing emotions in humans has attracted a lot of researchers in past two decades. The major challenge in this area is to recognize the features to be extracted from the speech that can effectively and efficiently classify the emotions in humans. Audio Emotion Analysis includes the detection of a scream, an extreme emotion of fear and classification of emotions. Scream detection is done using 2 methods. One involving extraction of log energy, auto-correlation and MFCC features from the input speech and other involving the use of non-extensive entropy on MFCC features. For audio emotion analysis, two methods are proposed. First method makes use of weighted mean and entropy of MFCC features and second method uses different entropy values of different audio features like pitch and energy. Speech database used are freely available for research purpose. SAVEE database is used for emotion recognition. For scream detection, many sounds like applause, laugh, different type of screams etc are taken from the web directly.