Cultural Broadcast Archive

Development of an audio classification software for the radio programmes in the Cultural Broadcasting Archive.

We aim at automatic annotation of music versus speech in the > 17.000 radio programmes in the online archive CBA (Cultural Broadcasting Archive). In literature, promising approaches for this task exist. These are mostly based on classical machine learning approaches with e.g. SVM classification, and features like MFCC. Seyrlehner et al. introduce the CFA (continuous frequency activiation), which integrates feature and classification and outputs a single numerical value for each chunk of audio.
CBA contents are mainly provided by community radios. In contrast to archives of professional broadcasters, the CBA contains a high variety of different audio qualities which result from different equipment qualities and different knowledges about audio engineering. Our challenge is to create a classification that is robust against this varying quality. We assume, that the combination of different features and diverse pre- and post processing steps will help to solve this task.

Project members 
Matthias Husinsky

Project partners

Publications

Wieser, E., M. Husinsky, and M. Seidl, "Speech/Music Discrimination in a Large Database of Radio Broadcasts from the Wild", Conference Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Florence, IEEE, pp. 2134-2138, 05/2014.