|Second Workshop on Mathematical Analysis of Images in Bordeaux
Srđan Kitić: Audio inpainting by sparse regularization methods.
Audio inpainting is a challenging inverse problem arising in many practical applications (as a consequence of impulsive noise or packet loss, for instance). We address a particular case of audio inpainting scenarios, namely de-clipping (i.e. de-saturation) inverse problem which can be seen as an extreme case of dynamic range compression. Clipping is a very common problem in audio recordings, either due to poor dynamic range of the microphones, or due to effects of analog-to-digital conversion.
Rich theory of sparse representations allows us to regularize this ill-posed inverse problem and enhance the corrupted signal. The recovery methods enforce the recovered signals to be consistent with the clipped observations which yields the high quality estimate, both in terms of signal-to-degradation ratio and acoustically. We show that the declipping problem can be solved blindly, without an a priori information of the saturation threshold, which is the most common case in real-life applications.