Improving the Reliability of Music Genre Classification using Rejection and Verification

This paper presents a novel approach for post-processing the music genre hypotheses generated by a baseline classifier. Given a music piece, the baseline classifier produces a ranked list of the N best hypotheses consisting of music genre labels and recognition scores. A rejection strategy is then applied to either reject or accept the output of the baseline classifier. Some of the rejected instances are handled by a verification stage which extracts visual features from the spectrogram of the music signal and employs binary support vector machine classifiers to disambiguate between confusing classes. The rejection and verification approach has improved the reliability in classifying music genres. Our approach is described in detail and the experimental results on a benchmark dataset are presented.