A Causal Rhythm Grouping

This paper presents a method to identify segment boundaries in music. The method is based on a multi-step model; first a features is measured from the audio, then a measure of rhythm is calculated from the feature, the diagonal of a self-similarity matrix is calculated, and finally the segment boundaries are found on a smoothed novelty measure, calculated from the self-similarity matrix. All the steps of the model have been accompanied with an informal evaluation, and the final system is tested on a variety of rhythmic songs with good results. The paper introduces a new feature that is shown to work significantly better than previously used features, a robust rhythm model and a robust, relatively cheap method to identify structure from the novelty measure.