Forensic speaker profiling in a Hungarian speech corpus
András Beke · 2018
Humans and also machines recognize the known persons based on their voice characteristics. If the speakers are unknown, some specific information or estimates may be still available about them: their sex, age, height, weight etc. Previous research has proven that there is a high correlation between the length of vocal tract and the speaker's physical conditions such as sex, age, height, weight etc. Results suggest that some acoustic- phonetic cues of human speech are related to body conditions. In this research we would like to examine how successfully can we predict the speaker's body conditions from their speech using machine learning techniques. We build various models to predict the speaker's sex, age, height and weight. All of the models are trained on collected speech features including prosody, voice quality and spectra based features. The contribution of the individual features is analysed. The results show that sex, body weight and height can be used to predict speaker characteristics with very high precision, however predicting the age is a more complex task i.e. predicting the age of older speakers gets more and more difficult. These results also allow for automatic speaker profiling in cognitive infocommunication or is forensic speech research.