Fast and Accurate: Improving a Simple Beat Tracker with a Selectively-Applied Deep Beat Identification

In music applications, audio beat tracking is a central component that requires both speed and accuracy, but a fast beat tracker typically has many beat phase errors, while an accurate one typically requires more computation. This paper achieves a fast tracking speed and a low beat phase error by applying a slow but accurate beat phase detector at only the most informative spots in a given song, and interpolating the rest by a fast tatum-level tracker. We present (1) a framework for selecting a small subset of the tatum indices that information-theoretically best describes the beat phases of the song, (2) a fast HMM-based beat tracker for tatum tracking, and (3) an accurate but slow beat detector using a deep neural network (DNN). The evaluations demonstrate that the proposed DNN beat phase detection halves the beat phase error of the HMM-based tracker and enables a 98% decrease in the required number of DNN invocations without dropping the accuracy.