Performance Analysis of ML Algorithms on Speech Emotion Recognition

Even though human–computer interface (HCI) applications such as computer-aided tutoring, learning and medical assistance have brought much changes in human lifestyle. This work has mainly focused on comparison of performance of five commonly used classifiers on emotion recognition. Since features are usually high-dimensional and structurally complex, the efficient classification has become more challenging particularly on low-cost processor and mobile (Android) environment. In this work, five machine-learning algorithms are implemented for speaker-independent emotion recognition and their performance is compared: (a) Logistic regression (LR), (b) K-nearest neighbour (KNN), (c) Naive Bayesian classifier (B), (d) Support vector machine (SVM), and (e) Multilayer perceptron (MLP) of neural network. The feature extraction techniques used to obtain features from speech are (a) Mel-scaled power spectrum; (b) Mel frequency cepstral coefficients. Naive Bayes classifier shows best results in speech emotion classification among other classifiers. Emotion data of happy and sad is taken from Surrey Audio-Visual Expressed Emotion (SAVEE) database.

Read in-depth Research Analysis