Exploring robust spectral features for emotion recognition using statistical approaches

Emotion recognition using speech is a natural way to identify the emotions of humans easily. Here the research work explored on different spectral features such as MFCC, pitch chroma, skewness, and centroid for emotion recognition. For this experimental classification, the emotions considered in this study are Fear, Anger, Neutral, and Happy. The motive of this work is to consider the male speech data and evaluating for various combinations of spectral features then finds the robust features. The system is finally it is figuring out that the combination of MFCC and skewness gave a better recognition performance when compared to other combinations. The above-mentioned features are examined using Gaussian Mixture models (GMMs) and Support Vector Machine (SVMs). To improve the system performance and avoid the high-dimensional data from the newly generated robust features, this work used an approach, namely called Principal Component Analysis (PCA) is for good recognition rate. It was established that the recognition performance for feature sets after applying PCA got increased in both classification models using GMMs and SVMs. The overall system is recognized 40% before PCA, 95% after PCA using GMMs and 31.6% before PCA, 75.8% after PCA using SVMs. The database used in this study is a Telugu emotion speech corpus (IIT-KGP).