Section: New Results
Audio segmentation and classification
This work has taken place in the context of the QUAERO Project.
New developments were given to audio segmenting and clustering tasks, w.r.t to new merging and stopping criteria for the clustering process. The main goal was to enhance the robustness of the BIC approach and to get a more reliable stopping criterion, thus leading to a better segmentation of audio documents.
We also developed and evaluated speech and music detection and classification algorithms on the QUAERO audio corpus (80 hours, among which 40 hours were annotated at IRISA). Test were carried out with several configurations of the AudioSeg toolkit and in various combinations of blind source separation methods.
Our best system performed at a state-of-the-art level. Source separation approaches did occasionnaly improve the results but they did not show yet a systematic advantage.