Nonparametric estimation of the dynamic range of music signals

摘要

The dynamic range is an important parameter which measures the spread of sound power, and for music signals it is a measure of recording quality. There are various descriptive measures of sound power, none of which has strong statistical foundations. We start from a nonparametric model for sound waves where an additive stochastic term has the role of catching transient energy. This component is recovered by a simple rate-optimal kernel estimator that requires a single data-driven tuning parameter. The distribution of its variance is approximated by a consistent random subsampling method that is able to cope with the massive size of the typical dataset. Based on the latter, we propose a statistic, and an estimation method that is able to represent the dynamic range concept consistently. The behaviour of the statistic is assessed based on a large numerical experiment where we simulate dynamic compression on a selection of real-world music signals. Application of the method to real data also shows how the proposed method can predict subjective experts' opinions about the hifi quality of a recording.