[vorbis-dev] idea for a new algorithm

[Monty]

> After the filterbank the maximum aplitude in each band decides
> about the nessesary bits, then the psychoacoustic model decides how
> much of the signal can be truncated without hearable noise.

Vorbis does not use critical bands or any subbanding.

> What do you think about quantising the remaining bits nonlinear
> with a logarithmic scale? This would share 25%.

That's entriely up to the choice of codebook, and the codebook can be whatever 
the encoder decides.

Ah, you mean quantizing time domain residual?  There are hooks for time domain
coding, yes, and Greg Maxwell has played with wavelet encoding this.  However,
this is not meant to catch inaudible residuals as much as it is to eliminate
impulses and plosions that do not represent well in the frequency domain.

> The Result would be a better resolution on low amplitudes.

There are not really absolute amplitudes of any sort involved... Vorbis has an
(obviously absurdly large) representation range of over 300dB.  Encoding a low
dB signal is not different than encoding a high dB signal.  The psychoacoustics
are designed to look at the range of the input signal, not how loud it would be
from an arbitrary zero dB point.

Monty

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/