What would it sound like if John Lennon sang any song written in the last 20 years?

I'll get back to that. First, here's how I'm a clever bastard.

We had our first meeting for the class I'm TAing this morning. We set up some meeting times and shuffled a couple of due dates around. Later, in class, Professor K announced the due date change for the first assignment - except instead of saying it was due at noon, which is what we had decided on at the meeting, she said it was due at 3 PM, which is when we were supposed to meet to grade it.

I leaned over to Katie, one of my co-TAs. "She means noon, right? Because we're meeting at 3."

"Yeah..." Katie raises her hand. "Prof. K, I think it's due at noon."

"That's right," says Prof. K. "We're grading it at 3. It's due at noon."

The students all turn around to glare at Katie. "Sorry," she says.

Neil, the other TA, looks over at me, smiling. "Wimp," he jibes.

"Jerk," chuckles Katie.

I shrug.

So I had this crazy idea during Computer Music today. (For those keeping score, Music Major Girl seems to have given up on getting off the waitlist.) Indulge me while I geek for a minute.

I was reading recently about a project someone had done for their doctoral thesis wherein they made an image filter that could learn analogies. So, for instance, take an aerial shot of a landscape - a river in a forest. Make a corresponding picture with a different color for each type of terrain (say, blue for water, green for forest). Now, make a new picture using those colors, and run all three through the filter. What comes out is an aerial shot of a landscape matching the new picture you drew, and it actually looks pretty realistic.

The idea I had is this: take the raw lead vocal tracks from, say, the Beatles "1" album. Now, take a singer with a high degree of vocal control - someone who can sing very consistently. Have him sing the songs, matching Lennon's rhythm as closely as possible.

Now, digitize the two tracks in the same way. Looking at the data over a one-second window, feed it into a neural network, with the new tracks as input and the originals as output. Move the window up a fraction of a second for each training example.

Once the neural network is trained, have the new singer sing any song at all. Feed the track into the input of the network, and record the output. If all goes well, then what you have is John Lennon singing the same song.

There are some issues here - it's not obvious how to record the output, and there's no guarantee that there's enough information contained in the raw audio data to generalize; some preprocessing might have to be done. Also, it'd be a pretty huge network, and require a lot of processing power.

But ideally, if it worked, you could reproduce the sound of any singer singing any song. If it worked really well, it would even capture certain nuances of the singer's style. That would be really cool.

Okay. End of geeking. Sorry to bore you.