Note: All plots are committed to this repository, but not all source sound files or generated data due to their large size. See the setup section below if you want to reproduce the experiments.

The experiments described here mainly follow Juang et al (1982) for the implementation of the algorithm for codebook generation as well as in terms of corresponding evaluation based on distortion performance. A summary of key relevant background including LPC analysis and clustering can be found here. In our case, we are exploring this technique for clustering of whale sounds. The reported experiments are initial steps toward the more general goal of possibly conforming a basis for development of a whale vocalization classification at the sub-unit level informed by and including comparison with similar efforts in the community.

In short, the steps followed in each experiment below are:

• LPC analysis to generate training vectors for clustering from a given whale sound file. This generates a single "predictor" file containing all the LPC vectors for codebook training.

• Codebook generation, inspection, and evaluation. This includes plots of distortion, σ-ratio, and inertia as functions of codebook size; cell cardinality and distortion; and visualization of the clusters and centroids based on the first couple of reflection coefficients associated with the prediction coefficients.

• HBSe_20161221T010133 (38,265 training vectors).

• HBSe_20170128T231621 (77,034 training vectors). Includes visualization of cluster shapes.

• 6_good_songs (268,686 training vectors). Includes visualization of cluster shapes.

• MARS_20161221_000046_SongSession_16kHz_HPF5Hz (1,108,820 training vectors from a single 4.5-hour recording). Includes visualization of codeword assignments on selected intervals having regions that are acoustically and visually (via spectrogram) similar.

To facilitate reproducibility this repository has ecoz2 as a submodule (to point to a particular version).

$ git clone --recursive https://github.com/ecoz2/ecoz2-whale-cb.git
$ cd ecoz2-whale-cb
# Build ecoz2 executables:
$ cd ecoz2/
$ make
$ export PATH=`pwd`/_out/bin:$PATH
$ export PATH=`pwd`/src/py:$PATH
$ pip install pandas matplotlib 
$ cd ..

Juang, B-H., Wong, D.Y., Gray, A.H., "Distortion Performance of Vector Quantization for LPC Voice Coding," IEEE Trans. ASSP, Vol. 30, No. 2, April, 1982. DOI: 10.1109/TASSP.1982.1163866