goal: stop analysis and start writing...

Part 1: About my to-do

So, I must admit I am having trouble letting go of the analysis part of this paper. And I must admit I have a problem.  A scoping problem.  I tend to be overambitious, but not realize that am being overambitious. I need to stop doing things that no one has done before for small projects.

Because those things are not small projects.  They have lots of crevices and unknowns that one can get buried in. And I need to limit the amount of work I do for small projects.  My time is limited.  And no one can finish or do well on a LOT of huge projects.

Also, speaking of which, 2nd dinosaur prototype. I can send it to my friend, and he will finish it -- I mean, the outside parts... while later I will make a larynx that works for it and test the audio part of it... I'm thinking of doing that, but it is hard to let go.  However, the dinosaur project has been collaborative from the start, even though I have been at the head of it... But the 3d modeling has always been done by others, while I meet with them and ask them to do things and specify my requirements, etc. At this point, I'm def. giving my friend 2nd authorship on any paper... And it is the best thing for the project, I think, not for *it* to be on hiatus while I'm in Buenos Aires doing interactive tango thingies. Still, I'm hesitating.  Hard to give up my baby, even its not just MY baby.

Also, I need to find a graceful way to exit my leadership positions of AMESA and Lorkas.  I'm leaving next semester, and so even if I had the time for them I'd have to let them go.  But yes, also, I need to start jealously guarding my time.

Sooooo:
++send some important emails
++start my paper.  prose.  PROSE.
++stop with all the analysis and say the stuffz.

Part 2: Self-similarity and data analysis journeyz

So, what I have been working on for the last 2 or so weeks is coming up with self-similarity maps, figuring out how to interpret them, and coming up with ways of measuring similarity (distance?).  It has been a rather exciting process, actually...

So, although I read a few papers on self-similarity maps and how to read them, I really didn't have a clear intuitive feel for how to intuit structure from seeing them.  I kinda knew, but I didn't feel confident... I was lacking that er, gut certainty? I finally realized that needed to make and see my own examples.

So, what I started off doing was creating self-similarity for different random signals.  This was confusing, but it wasn't until I realized that I really need to do an honest-to-god averaging filter on the audio data in order to really get the amplitude envelope that things made sense (ie, I could predict what the map was).  And then, by god, the self-similarity map of white noise is completely BLACK.  Which is what SHOULD happen.  White noise, by definition doesn't have structure. But actually, the poisson distribution had some superficial similarities to some of the results I was getting (some!).  That needs to be explored further --> depending on time...

Anyways, so then I made my own audio file, very short, that was clearly structured.  And then I changed it, and saw how that changed the similarity map.  So enlightening!!! At each step I stopped to show Brent -- LOOK!! So this means THIS!!! Etc.

Then, from that, I made a function in matlab to simplify my workflows.  Then, I made a whole bunch of nice-looking maps, and started looking at them on top of each other photoshop, and eying structure.  That's when I realized that I was seeing structure EVERYWHERE. Even in unrelated files.   Human brains are like that!!

It was then that I realized for my own comfort, I was going to have to come up with a way to compare two similarity maps computationally.  I needed something fast, so after about 10 minutes of googling and thinking, I realized that I could use background subtraction, then add up the remaining pixels to get a measure of similarity.  I could even visualize the result.

This..... wasn't so helpful.  It showed that the maps that I judged as closer to each other just by eyeing as actually more similar, but the differences were small.  I did more researching, and decided to compared averaged regions instead.  This gave more of a spread in the results, but still showed a lot of clustering in the same area of similarity (although less so than just pixel by pixel).  I tried some program (algorithms unknown -- there was a paper, but I used the program to see if it was worth reading it  (I would only read if I had to cite it))  for perceptual similarity but this didn't outperform the pixel by pixel test.  So far, the averaging produced the greatest spread.

The problem was still this clustering --- similarity maps kinda all looked similar to each other no matter what. So now I am finding the similarity maps of unrelated audio to movement as a comparison. First, I'm comparing a Webern piece to everything.  It matched (relatively) horribly with everything except for the giselle movement, which also has a lot of black space.  This is the weakness of this particular algorithm -- even if the white lines / spaces are in completely different parts, if you have a lot of blackness in the similarity diagram the two maps (of structure) will show as pretty similar. But I will be testing with a shostakovich cello concerto & an Ace of Base song, so the average non-related number should average relatively high.  I still wish I knew a way of getting a p-value from this though...