This was an experiment that began with making an audio segment playlist. What that meant was given a couple of songs, the code would identify the important representative segments of the songs and make a new playlist from them.
I broke the given music into its detected onsets (using essentia) and made segments of varying bar lengths. These segments were then clustered together (mfcc features were used to measure L2 similarity). The representative clusters were then identified to make a summary of the song.
I thought this was a smart way to summarize a song, or atleast identify the pivotal segments within a song. With this motivation I chose the song “Lateralus – Tool” which is exciting for a long list of reasons (read fibonacci music), and decided to summarize the 9:24 minute song to smaller glorious excerpts that captured the most variable sections of the song.
github link : Audio Summarizer ipython notebook
Manaswi Mishra 