Accurate Tempo Estimation based on Recurrent Neural Networks and Resonating Comb Filters

In this paper we present a new tempo estimation algorithm which uses a bank of resonating comb filters to determine the dominant periodicity of a musical excerpt. Unlike existing (comb filter based) approaches, we do not use handcrafted features derived from the audio signal, but rather let a recurrent neural network learn an intermediate beat-level representation of the signal and use this information as input to the comb filter bank. While most approaches apply complex post-processing to the output of the comb filter bank like tracking multiple time scales, processing different accent bands, modelling metrical relations, categorising the excerpts into slow/ fast or any other advanced processing, we achieve state-of-the-art performance on nine of ten datasets by simply reporting the highest resonator’s histogram peak.