Windowing for Speech Emotion Recognition

In this paper we present window time length analysis for speech emotion recognition. We performed our tests on spectrograms calculated from Hamming window of different length in time and trained our Convolutional Neural Network (CNN) architecture with two convolutional layers and one fully-connect layer. The goal was to find out which resolution of spectrograms, whether time or frequency one is better to classify speech signal into seven emotional classes successfully. According to our results, right choice of window length is important for Speech Emotion Recognition (SER). Accuracy depends on window length and overlap of consecutive frames. The results show that preprocessing of speech signal in time domain plays important role for SER system, even with usage of CNN as classifier.