Speech/music discrimination in a large database of radio broadcasts from the wild

This paper describes the development, implementation and evaluation of a speech/music detector. We aim at audio from different sources with different qualities - i.e. audio from ”the wild”. We examine existing approaches for audio classification and select a recent feature. We modify the feature and evaluate the classification accuracy on a random test set of more than 60 hours of audio material against a standard speech/music detection approach. With our approach, we reach a classification accuracy of 96,6%. We provide a performant open source implementation of our detector.