Beat Tracking: Is 441 kHz Really Needed?

Abstract

Beat tracking is essential in music informationretrieval, with applications ranging from music analysis and automaticplaylist generation to beat-synchronized effects. In recentyears, deep learning methods, usually inspired by well-knownarchitectures, outperformed other beat tracking algorithms. Thecurrent state-of-the-art offline beat tracking systems utilize temporalconvolutional and recurrent networks. Most systems use aninput sampling rate of 44.1 kHz. In this paper, we retrain multipleversions of state-of-the-art temporal convolutional networks withdifferent input sampling rates while keeping the time resolutionby changing the frame size parameter. Furthermore, we evaluateall models using standard metrics. As the main contribution,we show that decreasing the input audio recording samplingfrequency up to 5 kHz preserves most of the accuracy, and insome cases, even slightly outperforms the standard approach.