Hi Echonest team,
I am currently using the Analyzer API to extract timbre vectors from audio as part of my final year thesis in which I explore the potential of timbre dissimilarity as a source of information about difference in perceptual quality between tracks. I am using Echonest because I found some evidence that it compares well to other extraction algorithms.
I was able to find some really useful information about the algorithm behind the Analyzer by putting together info from Tristan's MIT thesis, the Analyzer documentation and different threads from this forum but I am still finding it hard to understand what is done to extract the timbre vectors. Also Tristan mentions in a post that the algorithm has been changed since the time of writing the thesis to a method that now includes PCA as the final stage of process. I understand the meaning of the PCA representations but I was not able to find any other info about how the PCA is applied to the signal (or what type of PCA) or how the confidence values are obtained, so I was just wondering if any other information is available from you.
Thanks a lot in advance.
Best,
Gianni