This thesis proposes blind audio source separation algorithms using a microphone array. The final application of these algorithms is robot audition through the ROMEO project. In this thesis, we developed blind source separation algorithms based on a sparcity criterion. We show that l1 minimization using the natural gradient optimization technique has the same performance that the state of the art. We show that a criterion based on the parametrization of the quazi-norm lp, with 0<p<1, improves the previous results: the sparsity criterion gets harder through the iterations of the algorithm. Then, we exploited the multisensor aspect of our application (16 sensors are fixed in the head of the humanoid) and we proposed a separation algorithms with a fixed beamforming preprocessing step. In the robot audition case, the sensors are often placed on the head of the humanoid. To take into account the influence of the head in the near sound manifold, we built the beamforming filters using the head related transfer functions (HRTF) of the robot. The beamforming step improves the separation results compared to the use of a blind source separation only. This thesis also proposes the adaptive versions of those algorithms. In the real scenario where the number of sources is unknown and changes, we show how to detect and follow the number of sources.
