Audio and Speech Classification Applied to Child Sexual Abuse Investigation

Abstract

The complexity and scale of seized media in criminal investigations has increased dramatically in recent times, not least in child sexual abuse investigations. Manual examination of material impose great stress on the investigator and innovative aids can play a crucial role mitigating this. The thesis evaluates the use of machine learning algorithms for automatic speech classification. More specifically, we present the components of a system that uses acoustic features to identify speech in noisy environments and classify the speakers gender and spoken language. For each of the tasks, separate approaches based on earlier research were developed and experiments were devised to validate them. The results of all classification tasks were satisfactory, but the language classifier were found not to scale well with the number of supported languages. In conclusion, the thesis shows that machine learning models are well suited for speech classification. The thesis was performed at Safer Society Group.