Extracting Sound Objects by Independent Subspace Analysis
In this paper we present a scheme for unsupervised extraction of sound objects or sources from a single recording containing a mixture of sounds. The separation/extraction procedure is performed by orthogonal projection of the mixed sound onto sub-spaces that are derived by clustering of transform coefficients, such as coefficients obtained by PCA or ICA. The clustering step reveals a residual non-linear grouping structure of the signal that is omitted by the linear transform. To achieve independence we are searching for partitioning that maximizes the mutual information between a component and a set to which it belongs. This information is obtained by considering a pairwise distance measure among all coefficients. Source separation experiments are reported in the paper.
Click to purchase paper or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members, $5 for AES members and is free for E-Library subscribers.