A large amount of modern pop music contains digital "loops" or "samples" (short audio clips) that appear multiple times during a song. In this paper, a novel approach to separating these exactly repeating component waveforms from the rest of an audio mixture is presented. By examining time-frequency representations of the mixture during several instances of a single repeating component and taking the complex value for each time-frequency bin with the smallest magnitude across all instances we can effectively extract the content which is perceived to be repeating given that the rest of the mixture varies sufficiently. Results are presented demonstrating successful application to commercially available recordings as well as to constructed audio mixtures achieving signal to interference ratios up to 42.6 dB.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.