An Entropy-Based Computational Classifier for Positive and Negative Emotions in Voice Signals

The detection, classification and analysis of emotions has been an intense research area in the last years. Most of the techniques applied for emotion recognition are those comprised by Artificial Intelligence, such as neural networks, machine learning and deep learning, which are focused on the training and learning of models. In this work, we propose a rather different approach to the problem of detection and classification of emotion within voice speech, regarding sound files as information sources in the context of Shannon’s information theory. By computing the entropy content of each audio, we find that emotion in speech can be classified into two subsets: positive and negative. To be able to perform the entropy computation, we first compute the Fourier transform to digital audio recordings, bearing in mind that the voice signal has a bandwidth 100 Hz and 4 kHz. The discrete Fourier spectrum is then used to set the alphabet and then the occurrence probabilities of each symbol (frequency) is used to compute the entropy for non-hysterical information sources. A dataset consisting of 1,440 voice audios performed by professional voice actors was analysed through this methodology, showing that in most cases, this simple approach is capable of performing the positive/negative emotion classification.