Tracking the odd: Bayesian meter inference in culturally diverse musical collections
In the field of MIR (Music Information Retrieval) the potential of Bayesian methods for signal analysis is widely acknowledged, but for many tasks in MIR Bayesian approaches are still the exception. In this talk a new state-of-the-art particle filter (PF) system is presented for the task of meter inference from a music audio signal. The system can be applied to determine the type of meter of a musical audio signal, and to track beats and downbeats in a unified framework. The new inference method is designed to overcome the problem of PFs in multi-modal state-spaces, which arise due to tempo and phase ambiguities in musical rhythm representations. Our results on a culturally diverse collection will be illustrated, in which the proposed model is shown to infer the type of meter in a collection of music samples from India, Greece, and Turkey. Our results suggest that the proposed system is capable of meter inference in large culturally diverse music collections. We argue that the system can be easily adapted to musical styles, and therefore avoids inclusion of an ethnocentric bias into music recommendation and distribution software systems.
presentation slides available here