Causal Prediction of Continuous-Valued Music Features

This paper investigates techniques for predicting sequences of continuous-valued feature vectors extracted from musical audio. In particular, we consider prediction of beatsynchronousMel-frequency cepstral coefficients and chroma features in a causal setting, where features are predicted as they unfold in time. The methods studied comprise autoregressive models, N-gram models incorporating a smoothing scheme, and a novel technique based on repetition detection using a self-distance matrix. Furthermore, we propose a method for combining predictors, which relies on a running estimate of the error variance of the predictors to inform a linear weighting of the predictor outputs. Results indicate that incorporating information on long-term structure improves the prediction performance for continuous-valued, sequential musical data. For the Beatles data set, combining the proposed self-distance based predictor with both N-gram and autoregressive methods results in an average of 13% improvement compared to a linear predictive baseline.