Algorithmic audio signal separation for sound effect applications

OData support
Szemenyei Márton
Department of Control Engineering and Information Technology

In my current work I address the problem of separating the sound signal components generated by each string in an output signal of an electric guitar.

In the literature of sound source separation several approaches can be found for the solution of a similar problem called the cocktail party effect with varying success. During my work I investigate the potentially applicable algorithms for the string separation problem with special respect to the so called Nonnegative Matrix Factorization method. I propose a system based on the usage of the algorithm called Sparse NMF with parameters optimized to solve the given problem.

To apply the supervised, trained method it is necessary to make a database that can store the characteristics of the sources. After the preparation of the database and the implementation of the SNMF method I describe several tests to evaluate the performance of the system based on the proposed method with respect to the string separation problem.

After the evaluation of the tests I made the conclusion that besides the system can produce quite good results on less complex given inputs, the robustness and the performance of the system needs to be improved either with further optimization of the NMF based method or with the application of a totally different algorithm.


Please sign in to download the files of this thesis.