Confidence Measures and Their Applications in Music Labelling Systems Based on Hidden Markov Models

Abstract

Inspired by previous work on confidence measures for tempo estimation in loops, we explore ways to add confidence measures to other music labelling tasks. We start by reflecting on the reasons why the work on loops was successful and argue that it is an example of the ideal scenario in which it is possible to define a confidence measure independently of the estimation algorithm. This requires additional domain knowledge not used by the estimation algorithm, which is rarely available. Therefore we move our focus to defining confidence measures for hidden Markov models, a technique used in multiple music information retrieval systems and beyond. We propose two measures that are oblivious to the specific labelling task, trading off performance for computational requirements. They are experimentally validated by means of a chord estimation task. Finally, we have a look at alternative uses of confidence measures, besides those applications that require a high precision rather than a high recall, such as most query retrievals.