Block-sparse Basis Sets for Improved Audio Content Estimation

Unsupervised lexicon learning techniques for audio-in-the-wild typically assume that only one of the lexical units is active at any given point in time (hard quantization) or use soft counts to avoid committing to one unit (soft quantization). In reality, the audio will usually be produced as a mixture of the different audio concepts in the lexicon. In this paper, we propose a model where the audio content is assumed to be generated by a mixture of a sparse subset of the lexical units thus guiding the system toward a better estimate of presence of the concepts. We present an approach that builds on current lexicon learning frameworks, and develop a novel algorithm to estimate the contribution of different sources by imposing block-sparsity constraints on the lexicon. Our proposed framework shows significant improvement over the standard lexicon learning framework on a retrieval task for audio-in-the-wild.