A comparison of Conditional Random Fields models for rhythm analysis
In this talk I will describe applications of Conditional Random Fields (CRFs) to rhythm analysis. CRFs are a framework for building probabilistic models for structured input–structured output data prediction, which have proven successful in a variety of real-world classification tasks. They can be used in several ways for the task of rhythm analysis. I will present and compare two recent approaches based on CRFs, one for the task of beat tracking, using a highly constrained temporal model and very simple features, and the other for the task of downbeat tracking, using deep learned features and mild temporal modelling. The main differences between the two models will be highlighted, as well as their strengths and limitations.