The effects of normalisation methods on speech emotion recognition

Abstract

Speech emotion recognition systems require features to be extracted from the speech signal. These features include Time, Frequency, and Cepstral-domain features. To normalise features, it is a challenging task to select an appropriate normalisation algorithm since the algorithm may impact classification accuracy. This paper presents the effects of different normalisation methods applied to speech features for speech emotion recognition. Speech features are extracted from the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) dataset and normalised before training machine and deep learning algorithms such as Logistic Regression, Support Vector Machine, Multilayer Perceptron, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). The CNNs and LSTMs obtained 72% for both accuracy and F1score outperforming standard machine learning algorithms. Feature normalisation improved both accuracy and F1score by more than 14% using CNN and LSTM.