5 ms·

It is my understanding that most voice codecs are still essentially lpc based, it is just that lpcs are unstable when quantised, so they get converted to things

It is my understanding that most voice codecs are still essentially lpc based, it is just that lpcs are unstable when quantised, so they get converted to things like line spectral pairs, then there is all sorts of different excitations and codebooks used to eek out every bit of savings. But the core idea of lpcs are still in there. Happy to be corrected by someone who knows better!

kragen 6y ago
I would be too! I think "LPCs are unstable when quantised" is a little overgeneralized — a more precise statement is that linear predictors, being all-pole filters, can be stable or unstable, and that quantization can convert a marginally stable or even absolutely stable filter into an unstable one. But I could easily be misunderstanding something basic! I've never understood the LSP thing. (Yet!)
hprotagonist 6y ago
it’s still LPC-ish, yes. a lot of the work in modern codecs goes into good VQ for the codebook, and better bit allocations. LPC10 sounds like ass, but damn is it cheap on bits!
- kragen 6y ago
  Mmm, ass! My favorite! The RPE in RPE-LTP — which is also one of the two excitations in LPC-10 — is indeed fairly precisely a fart sound, although the usual euphemism is "quasi-periodic pulses". That's because your larynx sounds like a fart if it doesn't have a vocal tract glommed onto the end of it to filter it with its recursive linear resonances. Fortunately sox has LPC-10 built in, so you can enjoy the tasty tasty 80s ass quite simply if you install sox: sox foo.wav foo.lpc10 play foo.lpc10 The LPC-10 code in Sox says it was translated from Fortran with f2c sometime around 1998, and as far as I can tell, nobody's been changing it since then. That's a nice counterpoint to the fucking "this web page doesn't work because nobody has changed it in the last five years" WWW. https://web.archive.org/web/20200420003350/https://www.jwz.org/doc/cadt.html https://web.archive.org/web/20200420003350/https://www.jwz.o...
  - hprotagonist 6y ago
    ah good stuff. I studied with one of Rabiner’s old coworkers, so this is a fun trip.

kragen 6y ago

I would be too! I think "LPCs are unstable when quantised" is a little overgeneralized — a more precise statement is that linear predictors, being all-pole filters, can be stable or unstable, and that quantization can convert a marginally stable or even absolutely stable filter into an unstable one. But I could easily be misunderstanding something basic! I've never understood the LSP thing. (Yet!)

hprotagonist 6y ago

it’s still LPC-ish, yes. a lot of the work in modern codecs goes into good VQ for the codebook, and better bit allocations. LPC10 sounds like ass, but damn is it cheap on bits!

kragen 6y ago
Mmm, ass! My favorite! The RPE in RPE-LTP — which is also one of the two excitations in LPC-10 — is indeed fairly precisely a fart sound, although the usual euphemism is "quasi-periodic pulses". That's because your larynx sounds like a fart if it doesn't have a vocal tract glommed onto the end of it to filter it with its recursive linear resonances. Fortunately sox has LPC-10 built in, so you can enjoy the tasty tasty 80s ass quite simply if you install sox: sox foo.wav foo.lpc10 play foo.lpc10 The LPC-10 code in Sox says it was translated from Fortran with f2c sometime around 1998, and as far as I can tell, nobody's been changing it since then. That's a nice counterpoint to the fucking "this web page doesn't work because nobody has changed it in the last five years" WWW. https://web.archive.org/web/20200420003350/https://www.jwz.org/doc/cadt.html https://web.archive.org/web/20200420003350/https://www.jwz.o...
- hprotagonist 6y ago
  ah good stuff. I studied with one of Rabiner’s old coworkers, so this is a fun trip.