When the pitch flag gets corrupted, make sure that pitch is actually
allowed instead of generating a bogus pitch index and crashing.

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

Change ec_{enc|dec}_bit_prob to take probabilities in units of 1/65536 instead of 1/256. This allows them to use a single 16x16->32 multiply instead of a 24x8->32 multiply. Also change the time-frequency resolution flag coding to ensure that "0" is always the most-probable symbol (i.e., that prob("1")<50%), as that's where all the rounding error accumulates.

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

Provide direct implementations ec_{enc|dec}_bit_prob() that do not require a division instead of using the normal entropy coder path. This should be exactly equivalent to the existing code.

Trying to fold only from regions where we have at least 2 bits of depth

Discarding the amplitude of the overlap for "frame_max"

Encoding the optimal tf-tradeoff for each band and then
applying it during quantization.

In cases where the coarse energy quantizer wasn't able to
go down fast enough, we could have a negative error that didn't
get clamped.

(rather than the lowest)

Apply a one-level Haar transform (or Hadamard) on each short MDCT
to nearly double the time-domain resolution.

When a transient occurs in the overlap between two frames, we
now make sure that both these frames are coded as transients

Adding one more level of band splitting so that splitting at the
PVQ encoding level is no longer necessary.

Also a bunch of comments

Now everything is done through quant_all_bands() and quant_band().

We prevent as many as possible, but they can still occur if we
manage to reach the LN limit

Uses a triangular PDF for coding the angle.