Dictionary learning methods for single-channel source separation.

Summary In this thesis, we propose three main contributions to dictionary learning methods. The first one is a group sparsity criterion adapted to NMF when the chosen distortion measure is the Itakura-Saito divergence. In most music signals one can find long intervals where only one source is active (solos). The group sparsity criterion that we propose allows to automatically find such segments and to learn a dictionary adapted to each source. These dictionaries are then used to perform the separation task in intervals where the sources are mixed. These two tasks of identification and separation are performed simultaneously in a single pass of our proposed algorithm. Our second contribution is an online algorithm for learning the dictionary on a large scale, over signals of several hours. The memory space required by an online estimated NMF is constant while it grows linearly with the size of the signals provided in the standard version, which is impractical for signals longer than one hour. Our third contribution concerns the interaction with the user. For short signals, blind learning is particularly difficult, and the provision of information specific to the processed signal is essential. Our contribution is similar to inpainting and allows to take into account time-frequency annotations. It is based on the observation that almost the entire spectrogram can be divided into regions specifically assigned to each source. We describe an extension of NMF to take into account this information and discuss the possibility of inferring this information automatically with simple statistical learning tools.