Tempo Octave Correction Using Multiclass Support Vector Machine

摘要

A measure of rhythm, the tempo is vital in any music retrieval, analysis and synthesis task. Tempo estimation algorithms in some cases provide wrong estimate that are usually multiple (1/2,1/3,2,3) of the actual tempo. We present a method base+octv to correct such estimates (octave errors) by using an octave classifier. The basic algorithm involving spectral flux method of onset detection, autocorrelation for obtaining self similarity and cross correlation with regular pulses to capture beat periodicity is inspired from previous works. Five features extracted from the audio are used to predict the octave error class which leads to corrected tempo. The evaluation of the technique is performed on an extensive dataset of 4011 audio files from popular datasets with a cross validation accuracy of 75.56%.