Make Yourself Transparent

by Adam Rokhsar

This song was composed by me in Protools and Max/MSP. I then performed chroma extraction in Matlab to obtain a pitch class profile by frame for the entire song. Though the octaves are lost, the data contains only an energy value for each of the 12 chromatic notes in Western music, which serves as a good template for identifying chords. The chroma features are considered the observation data for a hidden Markov model (HMM), and the hidden state is the chord which is responsible for the chroma distribution of that frame.

HMMs are a kind of finite state machine in which the probability of an observation is dependent on current state of the system. The state is not directly observable, or “hidden.” The model is comprised of three parameters: the initial state of the system (pi), the transitional probabilities between states (A), and the emission probability for each state (B).

Using the method described in Bello & Pickens (2005), I initialized the parameters of a hidden Markov model to reflect a priori knowledge of music theory. Therefore, A is initialized to a double-nested circle of fifth, since certain chord transitions are more likely than others (e.g, Cmaj to Fmaj, as opposed to Cmaj to F#maj); and B is initialized such that the mean and covariance of all the chroma values by state reflects the triad that makes up the chord. I then disallowed updating during training of B, and allowed A and pi to be updated as normal. The Baum-Welch algorithm was used during training, and the Viterbi algorithm yielded the most likely chord sequence to explain the chroma features.

Once the chord sequence was determined, I used audio-mosaicing to randomly replace each state-labeled frame with another frame from the song with the same label (Hoffman, Cook, & Blei, 2008). In other words, each frame that was labeled Cmaj is replaced with a different frame also labeled Cmaj. collage 8192 is the result.

References

Bello, J.P. and Pickens, J. A Robust Mid-level Representation for Harmonic Content in Music Signals. In Proceedings of the 6th International Conference on Music Information Retrieval (ISMIR-05), London, UK. September 2005.  

Hoffman, M., Cook, P., and Blei, D. Data-driven recomposition using the hierarchical Dirichlet process hidden Markov model.  In Proceedings of the 2008 International Computer Music Conference, Belfast, 2008.