G. Velikic and M.F. Bocko (USA)
audio processing, analytical signal, phase, source separation
Recognition of structure in an audio mixture is a complex task which may be simplified by first isolating the individual sources in the mixture. There are several approaches to audio source separation. In the fortunate event that one has as many mixtures available as there are independent sources the methods of independent component analysis may be employed. If only a single mixture of multiple sources is available then it is possible to rely on previous knowledge of the individual sources to accomplish the task, or alternatively, on events common to sets of overtones. We propose a source separation method that uses the mutual correlations of the phases of overtones to group information vertically, i.e., in the frequency dimension of a spectrographic sound representation. We use the cleanly separable overtones to create filters to isolate the remaining overtones that are too closely spaced in frequency to be separated by reasonable length band-pass filters. Finally, we use the history of the signal itself to group data horizontally (in time). We discuss the advantages and the limitations of this method using a number of illustrative examples and a proposed quantitative perceptual metric.
Important Links:
Go Back
