The goal of this study is to improve current state of the art in harmonic analysis from audio signals, developing a hierarchical model of relevance for jazz music styles. The first step, audio to chord label mapping, has been studied using chromagram features and various modeling strategies such as Hidden Markov Models (HMMs), Dynamic Bayesian Networks (DBNs), template matching and Deep Neural Networks (DNNs). We target to improve the state-of-the-art using chromagram features with DNNs architectures designed for chord estimation. Here, we also aim at testing the potential of incorporating instrument information (which is often available from meta-data) for improving efficiency of chord estimation. The next problem is the automatic parsing of chord sequences to map segments of the sequences to the tonal space. The recent advances in automatic analysis of harmonic structure starting from chord progressions showed that successful hierarchical structure models can be built using NLP. Another promising direction is the use of multi-dimensional vector representations and tree representations for chords progressions. An important work to be done is to develop automatic structural analysis methods for this jazz repertoire based on the harmonic analysis and also the development of ontologies to represent the theoretical background of jazz harmonies and with which we can better organize the corpus.

In this project there is also a relevant musicological, which is to perform comparative analysis of harmony in various subgenres of jazz music through analysis of chord progressions and tonal spaces. The subgenres in the first half of the 20th century (Ragtime, New Orleans, Swing, Bebop) will be considered since this will let us carry reproducible research where data can be shared without copyright limitations together with the tools developed. The first challenge is to define musicologically meaningful and domain (jazz) specific formal grammars, features for multiple view point representations and transformation (chord substitution) rules for chord progressions which can serve as the basis of representing and comparing progressions. Comparisons will be carried at various levels of a hierarchical model on a large number of recordings from these subgenres.