February 8th … another big deadline. Gemstone thesis presentation rehearsals are coming up and we are actually semi-prepared!

Our classifier is still staying true at around 70-75%. You can look at our latest trials at the following address:

http://psycho.pixelfaction.com/music/logs/

We shall be adding more and more logs as we do more trials.

At the last meeting, someone brought up the question of why we don’t just use spectrograms directly. Some of the answers thrown around included “it doesn’t contain enough information”. This is not entirely true. Spectrograms contain a LOT of data. Remember the 7280 number we throw around a lot? Well with spectrograms, this number turns into something like 333820 very very quickly. So if running our classifier 50 times over 150 songs takes about 4-5 hours, 333820 songs would take say… 230 hours, or about 9 days and some change (which comes out to an HOUR AND A HALF PER SONG). This is obviously not very ideal. Considering we don’t even know if it will work.

But why? You have 333820 points… it MUST be better! Well it could be, and we aim to find out by running it a couple of times, but even if it was, it would be hardly feasable for anything in real life. People don’t want to wait an hour and a half for a song to be classified on their music player (two minutes however, is fairly reasonable for something that is very much a work in progress).

We shall continue to try other classification routes now that our main classifier works fairly well. Stay tuned.

Cheers,
Aditya

Yes, they are.

But ours is doing a damn fine job. Our 3-way classifier, which I have named “ASH Classifier” (i.e. Angry-Sad-Happy Classifier). I know, I am a creative genius. In either case, the following is a confusion matrix for our classifier.

As you can see, our classifier is doing very well with happy and sad songs and quite decently with the angry songs. This is with a training set of about 40 songs per each class! Please keep an eye on the research section for more detailed explanation of our results.

We have preliminary results! The classifier is now classifying our test set of songs 80-85% correctly on average (thanks to the almighty Matlab master - Nima). This is distinguishing between happy and sad. Check out the research page for more details. For now, feast your eyes on some pretty figures that mean nothing to you!

Time-Averaged Rate vs. Scale for our happy songs

Time-Averaged Rate vs. Scale for our angry songs

Time-Averaged Rate vs. Scale for our sad songs

We also now have a photos section that is actually filled with photos. Enjoy.

Cheers,
The PsychoAcoustics Team

We came, we dressed nicely, and we conquered! That is all.

The IRB application for our survey is in! It was a bit of running around between the Lee and Ellicot buildings but hey, anything for the team?

In other news, I believe Aditya will get started on a private filemanager type area for the team.

Cheers.

Dancing ferrets and Power Rangers. That is all.

(read: watch this general space for updates)