PCA-Based Random Forest Classifier for Speech Emotion Recognition Using FFTF Features, Jitter, and Shimmer

In recent years, it has been observed that emotion recognition is a rapidly growing research area. Unlike humans, machines do not possess aptitudes to recognize emotions. This study aims to recognize five emotions, i.e., fear, anger, happiness, neutral, and sadness, from a speech signal. EmoDB dataset comprises the above mentioned emotions; so, this German dataset is used. Prosodic features like energy, voiced and unvoiced frame duration, and silence are extracted. When an individual expresses his/her emotions strongly, his/her voice amplitude and frequency fluctuations turn out to be high. Such fluctuations can detect an emotion, and it is possible to measure jitter and shimmer prosodic features that are extracted for a speech signal’s voiced frames. Spectrum tilt and zero-crossing rate are obtained, and frames of a speech segment are classified as voiced, unvoiced, and silence using hamming windows. Each frame comprises 960 speech samples. Moreover, in our study, a new feature, i.e., Fast Fourier transform (FFT), is extracted. Instead of using FFT of 960 samples, only mean, minimum, maximum, variance, and standard deviation for each FFT are used. We name these features as FFTF. Thirteen Mel frequency cepstral coefficients features are also extracted. A random forest classifier is used; in the first experiment, the effect of different number of trees is observed. With 100 trees, an emotion recognition accuracy of 78.6885% is attained. In the second experiment, different features of voiced and unvoiced frames are used, and it is observed that for FFT features, an accuracy of 78.6885% is attained. In the third experiment, the number of features are reduced using PCA, and it is observed that the accuracy is improved to 83.60%.