Tempo Induction Using Filterbank Analysis and Tonal Features

This paper presents an algorithm that extracts the tempo of a musical excerpt. The proposed system assumes a constant tempo and deals directly with the audio signal. A sliding window is applied to the signal and two feature classes are extracted. The first class is the log-energy of each band of a mel-scale triangular filterbank, a common feature vector used in various MIR applications. For the second class, a novel feature for the tempo induction task is presented; the strengths of the twelve western musical tones at all octaves are calculated for each audio frame, in a similar fashion with Pitch Class Profile. The timeevolving feature vectors are convolved with a bank of resonators, each resonator corresponding to a target tempo. Then the results of each feature class are combined to give the final output. The algorithm was evaluated on the popular ISMIR 2004 Tempo Induction Evaluation Exchange Dataset. Results demonstrate that the superposition of the different types of features enhance the performance of the algorithm, which is in the current state-of-the-art algorithms of the tempo induction task.