Event Synchronous Thumbnails: statistical properties

Large research efforts are devoted to the problem of music summarization and thumbnailing [4-5-6-7-8-9]. The application of these techniques are twofold, first application is to provide an efficient summary of the analyzed music in order to provide users with a first but short and efficient listening experience; second application is to provide automatic means to search in large music databases in order to identify music pieces. Our Event Synchronous Wavelet Thumbnails may be used for both purposes. They are based on the Event Synchronous Wavelet Transform [1], a recent evolution of the Multiplexed [3]and Pitch Synchronous Wavelet Transform [2] previously introduced by one of the authors; they have been recently proposed [1] in order to provide a comprehensive listening experience of a piece, a sort of timbral flavour or musical thumbnail; in this communication we will provide some results on the use of these thumbnails for the task of identification and search of music pieces in sound data bases. Our thumbnails are obtained by means of the computation of the Event Synchronous Wavelet Transform; a characterization of the thumbnail may be done by using the wavelet coefficient themselves or the MFCC or the LPC coefficient of the extracted Thumbnail, and will be a point in a multidimensional parameter space. Distance in this space or in a suitable projection of it will provide a measure of similarity and then a basis for search and identification. Some results in this direction confirm the starting hypothesis of the significance of the proposed thumbnails for identification purposes as well as for listening.