Chroma-based Predominant Melody and Bass Line Extraction from Music Audio Signals

In this dissertation we present the research work we have carried out on melody and bass line extraction from music audio signals using chroma features. First an introduction to the task at hand is given and important relevant concepts are defined. Next, the scientific background to our work is provided, including results obtained by state of the art melody and bass line extraction systems. We then present a new approach to melody and bass line extraction based on chroma features, making use of the Harmonic Pitch Class Profile (HPCP) [G´omez 06a]. Based on our proposed approach, several peak tracking algorithms for selecting the melody (or bass line) pitch classes are presented. Next, the evaluation methodology and music collections and metrics used for evaluation are discussed, followed by the evaluation results. The results show that as a salience function our proposed HPCP based approach has comparable performance to that of other state of the art systems, in some cases outperforming them. The tracking procedures suggested are shown to require further work in order to achieve a significant improvement in the results. We present some initial experiments on similarity computation, the results of which are very encouraging, suggesting that the extracted representations could be useful in the context of similarity based applications. The dissertation is concluded with an overview of the work done and goals achieved, issues which require further work and some proposals for future investigation.