A Study on Speaker Independent Emotion Recognition from Speech Signals

Abstract

In this work Spectral feature MFCC is considered for high accuracy, for Prosody feature Pitch is considered and concatenating with other two different features Cepstrum and DWT and this combination is called as Emotion- Specific Feature Set. A remarkable study is being done in the current years for improving human machine interaction on Speech Emotion Recognition. Speech contains a wealth of details about the Speaker's age, gender, and emotional condition.The challenging task of recognising a single emotion from a speaker is known as emotion recognition. The database under consideration is Telugu-Database, which is prompted by two male and female speakers and includes four emotions: happy, angry, sad, and neutral. To identify the corresponding emotion, different combinations of features are used, and these features are referred to as Emotion-specific features. The rate of combination identification is improved when these features are taken into account. The function information is extracted using the features DWT, Cepstrum, MFCC, and Pitch. After feature extraction, a back-propagation neural network algorithm is used to classify the data, and the results are then evaluated. The study concluded that by increasing number of nodes in the network and number of iterations the recognition rate is above 90%, the combinations of feature sets will give better emotion recognition rate than individual feature sets and the feature set combination DWT+Pitch+Cepstrum produced the individual emotion recognition rate above 95%.