A supervised learning method for tempo estimation of musical audio

Abstract

Automatic tempo estimation for musical audio with low pulse clarity presents challenges. In order to increase the pulse clarity of the input audio signals, the proposed method applies source filtering, especially low pass filtering, to the raw audio, so there are multiple audio clips for the processes. These processes are based on tempogram derived from onset detection function to obtain the tempo pair, which is the output of tempo-pair estimator, and their relative strength by the long-term periodicity (LTP) function. Finally, a classifier-based selector chooses the best estimated results from the different paths of audio. The performance of 1st place in at-least-one-tempo-correct index and 2nd place in P-score index in the evaluation MIREX 2013 audio tempo estimation demonstrate the effectiveness of the proposed method to audio tempo estimation.