Time-Frequency Distributions of Music based on Sparse Wavelet Packet Representations

We introduce a new method for generating time-frequency distributions, which is particularly useful for the analysis of music signals. The method presented here is based on ℓ1 sparse representations of music signals in a redundant wavelet packet dictionary. The representations are found using the minimization methods basis pursuit and best orthogonal basis. Visualizations of the time-frequency distribution are constructed based on a simplified energy distribution in the wavelet packet decomposition. The time-frequency distributions emphasizes structured musical content, including non-stationary content, by masking the energy from less structured music instruments. We present four examples for visualizing structured content, including vocal and single instrument