A Block Sparsity Approach to Multiple Dictionary Learning for Audio Modeling

Dictionary learning algorithms for audio modeling typically learn a dictionary such that each time frame of the given sound source is approximately equal to a linear combination of the dictionary elements. Since audio is non-stationary data, learning a single dictionary to explain all time frames of the sound source might not be the best modeling strategy. We therefore recently proposed a technique to jointly learn multiple dictionaries such that each time frame of the given sound source is approximately equal to a linear combination of the dictionary elements from one of the many dictionaries. This is equivalent to modeling each time frame with a small subset of all of the dictionary elements in the model, which is analogous to block sparsity on the mixture weights over all dictionary elements. In this paper, we show why there is inherent block sparsity in our model due to its hierarchical nature and why this is useful for audio applications.