Bayes GMM: Bayesian Gaussian Mixture Models

Python3 version

Overview

Both the finite Bayesian Gaussian mixture model (FBGMM) and infinite Gaussian mixture model (IGMM) are implemented using collapsed Gibbs sampling.

Examples and testing code

• Run make test to run unit tests.
• Run make test_coverage to check test coverage.
• Look at the examples in the examples/ directory.

Dependencies

• NumPy and SciPy: http://www.numpy.org/
• nose: https://nose.readthedocs.org/en/latest/

References and notes

If you use this code, please cite:

• H. Kamper, A. Jansen, S. King, and S. Goldwater, "Unsupervised lexical clustering of speech segments using fixed-dimensional acoustic embeddings", in Proceedings of the IEEE Spoken Language Technology Workshop (SLT), 2014.

In the code, references are made to the following:

• K. P. Murphy, "Conjugate Bayesian analysis of the Gaussian distribution," 2007, [Online]. Available: http://www.cs.ubc.ca/~murphyk/mypapers.html
• K. P. Murphy, Machine Learning: A Probabilistic Perspective. Cambridge, MA: MIT Press, 2012.
• F. Wood and M. J. Black, "A nonparametric Bayesian alternative to spike sorting," J. Neurosci. Methods, vol. 173, no. 1, pp. 1-12, 2012.

Some notes on the mathematical details can also be found at:

• http://www.kamperh.com/notes/kamper_bayesgmm13.pdf