Speech Emotion Recognition using a Novel Feature Vector based on Voiced Probability and Speech Characteristics

Emotion recognition in speech is one of the fast growing fields in artificial intelligence and machine learning. Feature selection is an essential process for speech emotion recognition and hence in this study a feature vector is proposed which is based on various characteristic of speech such as Zero Crossing rate, Fundamental Frequency and Mel Frequency Cepstrum Coefficients. The novelty that is introduced in this feature vector is using Probability of whether a frame is voiced or not This probability is concatenated with the rest of vector. The feature vector was tested on various classifiers such as CNN, SVM, Decision Tree, Logistic Regression. The result showed that the novel feature vector performed better as compared to conventional features used which is MFCC and its derivatives. Besides testing this feature vector against MFCC and its derivatives, feature vector (combined with a classifier) was also tested against other system of SER based on ML. The results showed that combination of feature vector and CNN were good and even outperformed some state of art techniques.