A Segmentation-based Tempo Induction Method

The automatized beat detection and localization have been the subject of multiple research in the field of music information retrieval. Most of the methods are based on onset detection. We propose an alternative approach: Our method is based on the “Forward-Backward segmentation”: the segments may be interpreted as attacks, decays, sustains and releases of notes. We process the segment boundaries as a weighted Dirac signal. Three methods devived from its spectral analysis are proposed to find a periodicity which corresponds to the tempo. The experiments are carried out on a corpus of 100 songs of the RWC database. The performances of our system on this base demonstrate a potential in the use of a “ ForwardBackward Segmentation” for temporal information retrieval in musical signals.