Timothy B. Terriberry authored 14 years ago and Jean-Marc Valin committed 14 years ago

Use a closed-form formula for the search instead.
This requires an integer sqrt, so it is not actually closed-form,
 but the number of iterations is O(qb) instead of O(2**qb).

Timothy B. Terriberry authored 14 years ago and Jean-Marc Valin committed 14 years ago

Use a closed-form formula for the codebook parameters instead.

Also convert the stereo split code to use log(N)/2 as the bit allocation
offset

Timothy B. Terriberry authored 14 years ago and Jean-Marc Valin committed 14 years ago

The old code allocated too many fine bits to large bands.
New allocations were derived from by numerical optimization using quantization
 MSE sampled from Laplacian distributed random data to within +/- 1 bit for
 N=2...160 and bits per band from 0 to 64.
Those allocations could be modeled with only minor errors using a simple offset
 of 19/8+log2(N), with no bits spent on fine energy when there would not be
 enough bits remaining to code a single pulse.
However, PEAQ testing suggested an offset of 14/8 was better, and that it was
 always worth spending at least one bit on fine energy.

Making the preemphasis memory an array of two, removing unused fields

Prediction coefficient now depends on the frame size. Also, the pdfs and means
have been adjusted and better quantized. This breaks compatibility of course.

a start band.

It is now the encoder's responsability to take care of it to avoid
busting the budget.

than what they're running at.

number of samples

as a denominator

Also dropped support for radices above 5.