Sentiments analysis in speech data

Details

Supervisors

Van Oirbeek, Robin ; Pennisi, Andrea

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

This master thesis explores sentiment analysis in speech data, focusing on its intricacies and the methodologies that can optimize its accuracy. Initial sections delve into the area of signal processing, studying sound properties and the application of transforms like the Fourier transform and its derivatives. A comprehensive literature review provides insights into the current state of the art and the advancements made in this domain. In the methodology section, emphasis is placed on the techniques of feature extraction pivotal for sentiment analysis. This involves an in-depth examination of feature selection and dimensionality reduction methods. The study evaluates different models, notably the Support Vector Machine (SVM), Multilayer Perceptron (MLP), and MLP regression, in their effectiveness for sentiment classification. The application of this research, utilizing features such as the first 20 Mel-frequency cepstral coefficients (MFCC), zero crossing rate, and root mean squared, revealed SVM to be particularly potent. When classifying sentiments into categories like "sad", "happy", "angry", and "calm", SVM achieved an accuracy of 85%, significantly outperforming the MLP which registered at 74%. This was tested on a mixed dataset combining RAVDESS and SAVEE dataset. The findings of this study underscore the potential of SVM in sentiment analysis of speech data and pave the way for further research to enhance accuracy and utility in real-world applications.