Audio Signal Separation Using Independent Subspace Analysis and Improved Subspace Grouping
Jens Wellhausen · 2006
Systems that perform the task of automatic search, retrieval or classification on audio signals are based on automatic data exploration algorithms. For most audio signals, blind source separation is an important preprocessing step before further classification becomes reliable or even possible. Blind source separation is a wide field of current research, and independent subspace analysis (ISA) seems to be promising to deal with much kinds of audio signals. The result of an ISA performed on audio signals is an over-separated set of subspaces. In this paper, new grouping algorithms for the recombination of the audio sources out of the separated subspaces are presented. After a review on ISA, different distance measurements for grouping algorithms are discussed in the first part. These distance measurements are used within three different grouping algorithms that are presented in the second part