Re: Linear prediction question
Posted: Mon Jun 04, 2012 1:30 pm
Update:
I've modified the format slightly, and now it not only sounds better, but has better compression (26.67% [frame size = 15] or 28.57% [frame size = 7]).
What I ended up doing was scrapping the 'multiple prediction models' idea and using a single model for all the data, and increase the order to 24 just because I can and it sounds best (and doesn't affect the actual sample compression at all - just increases the header size).
The compressed data is broken up into frames [no blocks, this time] of 7 or 15 samples [depending on how much quality you want].
There's a 4-bit 'shift' value [2^n scalar for...] and n nybbles [where n is 7 or 15]. This results in each frame fitting in 32 or 64 bits.
Of course, you can't really judge by just reading numbers so I've uploaded a few demonstrations here.
From what I've noticed, unpredictable samples [like bells, saw leads, etc] tend to sound worst, but other ones [like pianos, etc] sound best.
EDIT:
Oops, looks like I uploaded a wrong sample for the bells... heh. xP
I've modified the format slightly, and now it not only sounds better, but has better compression (26.67% [frame size = 15] or 28.57% [frame size = 7]).
What I ended up doing was scrapping the 'multiple prediction models' idea and using a single model for all the data, and increase the order to 24 just because I can and it sounds best (and doesn't affect the actual sample compression at all - just increases the header size).
The compressed data is broken up into frames [no blocks, this time] of 7 or 15 samples [depending on how much quality you want].
There's a 4-bit 'shift' value [2^n scalar for...] and n nybbles [where n is 7 or 15]. This results in each frame fitting in 32 or 64 bits.
Of course, you can't really judge by just reading numbers so I've uploaded a few demonstrations here.
From what I've noticed, unpredictable samples [like bells, saw leads, etc] tend to sound worst, but other ones [like pianos, etc] sound best.
EDIT:
Oops, looks like I uploaded a wrong sample for the bells... heh. xP