An efficient method for the unsupervised discovery of signalling motifs in large audio streams
Résumé
Providing effective tools to navigate and access through long audio archives, or monitor and classify broadcast streams, proves to be an extremely challenging task. Main issues originate from the varied nature of patterns of interest in a composite audio environment, the massive size of such databases, and the capability of performing when prior knowledge on audio content is scarce or absent. This paper proposes a computational architecture aimed at discovering occurrences of repeating patterns in audio streams by means of unsupervised learning. The targeted repetitions (or motifs) are called signalling, by analogy with a biological nomenclature, as referring to a broad class of audio patterns (as jingles, songs, advertisements, etc...) frequently occurring in broadcast audio. This paper evaluates the applicability of a pattern discovery architecture to a near-duplicate discovery task, that is the task of discovering and collecting occurrences of repeating audio patterns (also called motifs) bearing limited variability, in the absence of prior acoustic or linguistic cues. We adapt a system originally developed for word discovery applications, and demonstrate its effectiveness in a song discovery scenario. The adaption consists in speeding up critical parts of the computations, mostly based on audio feature coarsening, to deal with the large occurrence period of repeating songs in radio streams.