Look-ahead online NMF for unsupervised source separation

Algorithms based on Non-Negative Matrix Factorization (NMF) are commonly used to solve the Blind Source Separation (BSS) problem. The objective of NMF is to split a spectrogram, which is a frequency-time representation of a mixture, in two matrices; the basis matrix and the gain matrix. The most commonly variant used is the iterative NMF algorithm which requires the number of components as well as the full mixture spectrogram. The number of components can be approximated to the number of acoustical events presented in the mixture. Nevertheless, a spectrogram can be quite large, depending on the length of the mixture. A larger mixture introduces high latency and demands larger memory. The main problem is the number of components because it is not usually known beforehand. Therefore, Online NMF algorithms were developed to avoid these problems. The spectrogram of a mixture is divided in time segments called frames. Thus, instead of iterate over the full spectrogram, the basic online algorithm (ONMF) iterates over frames of the spectrogram to obtain the basis and the gain vectors. Nevertheless, if the basis vectors are estimated with just one frame of the spectrogram, some errors can be introduced due to the fact that a musical note usually requires more than one frame to be developed. In this thesis, the Look-ahead ONMF (LONMF) algorithm is proposed as an improvement of ONMF in order to obtain a better estimation of the basis matrix by taking more frames into account.Ingeniería de Telecomunicació