a novel algorithm to segment

an audio piece into its structural components.

The boundaries of the homogeneous regions are decided

based on various time and frequency domain

features. The algorithm has been designed in 2 stages.

In the first stage, a vocal/non-vocal/silence classification

is done using multinomial softmax regression. The

second stage uses a hidden Markov model to ‘smooth’

the previous output as well as enforce the time dependent

structuring.