home   people   research   events   teaching   vacancies   links

Gender recognition for improvement of aggression detection

Problem
Verbal aggression detection is based on certain voice properties, such as pitch and spectral shape. These values differ naturally between men, women and children. Therefore we would like to use different threshold values for these different groups. However, we do not in general have prior knowledge about the speaker.

Possible solution
Several techniques for gender recognition are known from literature. These are reported to have accuracies up to 100%. However, some systems use cues other than the sound alone and are therefore not useful. Other systems use cues based on sound representations as used in speech recognition systems, which differ from the cochleogram. All systems have only been tested on clean datasets, recorded under optimal conditions.

Proposal
We would like to adapt these techniques to work on the cochleogram, and then test them on both a clean dataset and a 'live dataset' comprised of recordings made by the aggression detection system. If proven to be useful, the techniques will be implemented in a future aggression detection system.

Contact
Dirkjan Krijnders ()