A Multi-model Approach to Beat Tracking Considering Heterogeneous Music Styles

In this paper we present a new beat tracking algorithm which extends an existing state-of-the-art system with a multi-model approach to represent different music styles. The system uses multiple recurrent neural networks, which are specialised on certain musical styles, to estimate possible beat positions. It chooses the model with the most appropriate beat activation function for the input signal and jointly models the tempo and phase of the beats from this activation function with a dynamic Bayesian network. We test our system on three big datasets of various styles and report performance gains of up to 27% over existing stateof-the-art methods. Under certain conditions the system is able to match even human tapping performance.