Illustrative Example 3 : Auto Segmentation

The Office Sounds data set (http://class-specific.com/os/) consists of 24 data classes of short-duration sounds, with 102 recordings of each sound class. The sounds are created by dropping objects of operating office equipment like staplers and scissors. Each recording is 0.5 seconds long. The sounds are consistently made, yet sometimes with just sublte difference between classes. It is an excellent dat set for testing classifiers for acoustic data.

The script software/officesounds.m is the upper-level script that implements the experiment by calling the subroutine software/mrhmm_officesounds.m once for each data holdout set. The script software/mrhmm_officesounds.m essentially follows the method of software/mrhmm_test1.m with automatic labeling.

With the script as provided, a performance of about 65 errors in 12240 chances (or 0.5% error) is achieved. An unpublished paper http://class-specific.com/csf/papers/eusipco_mrhmm.pdfgoes into more detail.