Using Hyper-genre Training to Explore Genre Information for Automatic Chord Estimation

Recently a large amount of new chord annotations have been made available. This raises hopes for further development in automatic chord estimation. While more data seems to imply better performance, a major challenge however, is the wide variety of genres covered by these new data. As a result, the genre-independent training scheme as is common today is bound to fail. In this paper we investigate various options for exploring genre information for chord estimation, while also maximally exploiting the full dataset. More specifically, we propose a hyper-genre training scheme in which each genre cluster has its own parameters, tied together by hyper parameters as a Bayesian prior. The results are promising, showing significant improvements over other prevailing training schemes.