Model-based cover song detection via threshold autoregressive forecasts

Current systems for cover song detection are based on a model-free approach: they basically search for similarities in descriptor time series reflecting the evolution of tonal information in a musical piece. In this contribution we propose the use of a model-based approach. In particular, we explore threshold autoregressive models and the concept of cross-prediction error, i.e. a measure of to which extent a model trained on one song's descriptor time series is able to predict the covers'. Results indicate that the considered approach can provide competitive accuracies while being considerably fast and with potentially less storage requirements. Furthermore, the approach is parameter-free from the user's perspective, what provides a robust and straightforward application of it.