When humans speak, the produced series of acoustic signs do not encode only the linguistic message they wish to communicate, but also several other types of information about themselves and their states that show glimpses of their personalities and can be apprehended by judgers. As there is nowadays a trend to film job candidate’s interviews, the aim of this Thesis is to explore possible correlations between speech features extracted from interviews and personality characteristics established by experts, and to try to predict in a candidate the Big Five personality traits: Conscientiousness, Agreeableness, Neuroticism, Openness to Experience and Extraversion. The features were extracted from a genuine database of 44 women video recordings acquired in 2020, and 78 in 2019 and before from a previous study.
Even though many significant correlations were found for each years’ dataset, lots of them were proven to be inconsistent through both studies. Only extraversion, and openness in a more limited way, showed a good number of clear correlations. Essentially, extraversion has been found to be related to the variation in the slope of the pitch (usually at the end of sentences), which indicates that a more "singing" voice could be associated with a higher score. In addition, spectral entropy and roll-off measurements have also been found to indicate that larger changes in the spectrum (which may also be related to more "singing" voices) could be associated with greater extraversion too.
Regarding predictive modelling algorithms, aimed to estimate personality traits from the speech features obtained for the study, results were observed to be very limited in terms of accuracy and RMSE, and also through scatter plots for regression models and confusion matrixes for classification evaluation. Nevertheless, various results encourage to believe that there are some predicting capabilities, and extraversion and openness also ended up being the most predictable personality traits. Better outcomes were achieved when predictions were performed based on one specific feature instead of all of them or a reduced group, as it was the case for openness when estimated through linear and logistic regression based on time over 90% of the variation range of the deltas from the entropy of the spectrum module. Extraversion too, as it correlates well with features relating variation in F0 decreasing slope and variations in the spectrum. For the predictions, several machine learning algorithms have been used, such as linear regression, logistic regression and random forests.