Undergraduate dissertation project exploring within-voice-variability using a voice classifier based on an artifical neural network (ANN).
Research on voice identity perception has extensively studied between-speaker variability in an attempt to understand how listeners tell people apart. Within-speaker variability, the way the same voice varies across situations, has mostly been eliminated in these studies, effectively treating it as a form of random noise that impairs voice identity perception. However, recent studies focusing on the ability to tell people together suggest that within-speaker variability is speaker-specific and plays an equally important role in identifying others. This study attempted to quantitatively verify these claims by training and testing a machine-learning model on variable and non-variable voice samples, examining emerging differences in identification accuracy based on differences in their training and test sets. Our results revealed that the complexity of variable voices makes them more difficult to identify compared to non-variable voices. However, the model was also able to extract speaker-specific information from variable voice samples that improved identification accuracy for variable voices. These findings show that within-speaker variability provides important cues for the identification of naturally varying voices, highlighting the need for further research on its role in voice identity perception, as well as demonstrating how the use of machine-learning models can benefit future psychological research.
In order to comply with ethics policies set by the University of Glasgow, this project does not contain any source data. Usage of this code in any form of educational assessment will most likely be detected by plagiarism software and should therefore be avoided.