Back to list

Separability of instruments in monophonic recordings

Authors Frauscher, E.
Year 2017
Thesis Type Audio Engineering project
Topic Audio Signal Processing
Keywords acoustics, audio recording and reproduction, Music Information Retrieval, signal processing
Abstract In this project thesis, the Non-negative-Matrix-Factorization (NMF) is chosen to be the method for source separation in monophonic audio signals. The approach is implemented and evaluated in Matlab. With the resulting toolbox it is possible to load audio signals and training data, choose the components which should be separated and analyse the result in a visual and aural way. For the initialisation different methods and datasets are tested. The factorization of the signal in the frequency domain is realised with the multiplicative update algorithm and variable cost functions. NMF only uses the absolute values to gain the spectral masks and thus causes some artefacts during the separation process. Therefore, additional optimizations like phase smoothing in the post-processing are implemented. Possible applications of this source separation are crosstalk-cancellation on stages or source separation for spatialization in 3D audio.
Supervisors Sontacchi, A.