A Post-Processing Procedure for Improving Music Tempo Estimates Using Supervised Learning

Tempo estimation is a fundamental problem in music information retrieval and has been researched extensively. One problem still unsolved is the tendency of tempo estimation algorithms to produce results that are wrong by a small number of known factors (so-called octave errors). We propose a method that uses supervised learning to predict such tempo estimation errors. In a post-processing step, these predictions can then be used to correct an algorithm’s tempo estimates. While being simple and relying only on a small number of features, our proposed method significantly increases accuracy for state-of-the-art tempo estimation methods.