| Repetition is a fundamental element in generating and perceiving structure. In audio, mixtures can often be understood as a background component that is generally more repeating in time, superimposed with a foreground component that is generally more variable in time (e.g., a song with varying vocals overlaid on a repeating accompaniment, or a recording with a varying speech mixed up with a repeating noise). On this basis, I present the REpeating Pattern Extraction Technique (REPET), a novel and intuitive approach for separating the repeating background from the non-repeating foreground in an audio mixture. The basic idea is to identify repeating elements in the mixture by measuring self-similarity along time, derive repeating models by averaging the repeating elements over their repetitions, and extract the repeating structure by comparing the repeating models to the mixture. Unlike other source separation approaches, REPET does not depend on special parametrizations, does not rely on complex frameworks, and does not require external information. Because it is only based on repetition, it has the advantage of being simple, fast, blind, and therefore completely and easily automatable. |
