Variability Tolerant Audio Motif Discovery

International audienceMining of repeating patterns is useful in inferring structure in streams and in multimedia indexing, as it allows to summarize even large archives by small sets of recurrent items. Techniques for their discovery are required to handle large data sets and tolerate a certain amount of variability among instances of the same underlying pattern (like spectral variability and temporal distortion). In this paper, early approaches and experiments are described for the retrieval of such variable patterns in audio, a task that we call audio motif discovery, for analogy with its counterpart in biology. The algorithm is based on a combination of ARGOS to segment the data and organize the search of the motifs, and a novel technique based on segmental dynamic time warping to detect similarities in the audio data. Moreover, precision-recall measures are defined for evaluation purposes and preliminary experiments on the word discovery case are discussed