![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
I hadn't stayed at work past midnight for quite a while. I was trying to work the bugs out of a FFT-based pitch shifter written in my thesis system, comparing it against a C benchmark, and trying to get it done for a Tuesday-morning meeting. It was a matter of dumping internal data from one stage of the system, fixing bugs until they matched, then moving to the next stage -- with all kinds of spurious mismatches due to roundoff feeding into floating-point mod and into arctan(0/0) situations. Then at the output mixing stage, I could not get my output to match his. Still can't. But I ran the C code (wrapped in my driver routine) and it generates output that blows up to 1.0e+33 and such, so the hell with that. I still don't see what's wrong with it.
So I came out of this still not knowing if my pitch-shifter is accurate. It sounds decent; heavily smeary/chorusy, but phase vocoder manipulation does have that artifact. (If you're bored, .wav files in http://www.cs.cmu.edu/~eli/tmp/t/ )
Sometime around 12:30 I had an uplifting realization: it doesn't matter if it's right. My thesis only claims that it can be done, because the system lets you build the data type to work directly with spectral data. A working pitch-shifter is not part of my thesis. Sometime very shortly after 12:30 I had a depressing meta-realization: so all this hard work is a waste of my time. Well, not totally; I do understand now how you translate between FFT-bin phase accumulation and accurate frequency, I think.
So today I didn't really accomplish any work, just decompressed and caught up on cruft and went shopping for gift books at Townsend and Caliban (Bryn Mawr is still closed). Unavoidably picked up some non-gift books, you know the way they adhere to your hair and clothing with miniature hooks when you brush against them?
So I came out of this still not knowing if my pitch-shifter is accurate. It sounds decent; heavily smeary/chorusy, but phase vocoder manipulation does have that artifact. (If you're bored, .wav files in http://www.cs.cmu.edu/~eli/tmp/t/ )
Sometime around 12:30 I had an uplifting realization: it doesn't matter if it's right. My thesis only claims that it can be done, because the system lets you build the data type to work directly with spectral data. A working pitch-shifter is not part of my thesis. Sometime very shortly after 12:30 I had a depressing meta-realization: so all this hard work is a waste of my time. Well, not totally; I do understand now how you translate between FFT-bin phase accumulation and accurate frequency, I think.
So today I didn't really accomplish any work, just decompressed and caught up on cruft and went shopping for gift books at Townsend and Caliban (Bryn Mawr is still closed). Unavoidably picked up some non-gift books, you know the way they adhere to your hair and clothing with miniature hooks when you brush against them?
