Toward a Sparse and Interpretable Audio Codec