Music tonality features for speech/music discrimination

We introduce a novel set of features for speech/music discrimination derived from chroma vectors, a feature that represents musical tonality. These features are shown to outperform other commonly used features in multiple conditions and corpora. Even when trained on mismatched data, the new features perform well on their own and also combine with existing features for further improvement. We report 97.1% precision on speech and 93.0% precision on music for the Broadcast News corpus using a simple classifier trained on a mismatched corpus.