This is an example of an accurate segmentation where the DTM model captures almost all the reference segments but incorrectly divides the verse (class 2) into 2 parts (these in fact correspond to singing / no singing
This is an example of a poor segmentation where the DTM model under-segments the ``refrain" class (class 3) and the constraints incorrectly expand class 6.
The addition of the musical constraints allows the DTM model to remove extra segment classes when there are more mixture components than necessary.
When there are more classes in the reference segmentation than there are DTM components, the model successfully ignores the smallest classes.
The addition of the musical constraints removes short segments.