This commit implements the perceptual noise substitution AAC extension. This is a proof of concept
implementation, and as such, is not enabled by default. This is the fourth revision of this patch,
made after some problems were noted out. Any changes made since the previous revisions have been indicated.
In order to extend the encoder to use an additional codebook, the array holding each codebook has been
modified with two additional entries - 13 for the NOISE_BT codebook and 12 which has a placeholder function.
The cost system was modified to skip the 12th entry using an array to map the input and outputs it has. It
also does not accept using the 13th codebook for any band which is not marked as containing noise, thereby
restricting its ability to arbitrarily choose it for bands. The use of arrays allows the system to be easily
extended to allow for intensity stereo encoding, which uses additional codebooks.
The 12th entry in the codebook function array points to a function which stops the execution of the program
by calling an assert with an always 'false' argument. It was pointed out in an email discussion with
Claudio Freire that having a 'NULL' entry can result in unexpected behaviour and could be used as
a security hole. There is no danger of this function being called during encoding due to the codebook maps introduced.
Another change from version 1 of the patch is the addition of an argument to the encoder, '-aac_pns' to
enable and disable the PNS. This currently defaults to disable the PNS, as it is experimental.
The switch will be removed in the future, when the algorithm to select noise bands has been improved.
The current algorithm simply compares the energy to the threshold (multiplied by a constant) to determine
noise, however the FFPsyBand structure contains other useful figures to determine which bands carry noise more accurately.
Some of the sample files provided triggered an assertion when the parameter to tune the threshold was set to
a value of '2.2'. Claudio Freire reported the problem's source could be in the range of the scalefactor
indices for noise and advised to measure the minimal index and clip anything above the maximum allowed
value. This has been implemented and all the files which used to trigger the asserion now encode without error.
The third revision of the problem also removes unneded variabes and comparisons. All of them were
redundant and were of little use for when the PNS implementation would be extended.
The fourth revision moved the clipping of the noise scalefactors outside the second loop of the two-loop
algorithm in order to prevent their redundant calculations. Also, freq_mult has been changed to a float
variable due to the fact that rounding errors can prove to be a problem at low frequencies.
Considerations were taken whether the entire expression could be evaluated inside the expression
, but in the end it was decided that it would be for the best if just the type of the variable were
to change. Claudio Freire reported the two problems. There is no change of functionality
(except for low sampling frequencies) so the spectral demonstrations at the end of this commit's message were not updated.
Finally, the way energy values are converted to scalefactor indices has changed since the first commit,
as per the suggestion of Claudio Freire. This may still have some drawbacks, but unlike the first commit
it works without having redundant offsets and outputs what the decoder expects to have, in terms of the
ranges of the scalefactor indices.
Some spectral comparisons: https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Original.png (original),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_NO.png (encoded without PNS),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS1.2.png (encoded with PNS, const = 1.2),
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/Difference1.png (spectral difference).
The constant is the value which multiplies the threshold when it gets compared to the energy, larger
values means more noise will be substituded by PNS values. Example when const = 2.2:
https://trac.ffmpeg.org/attachment/wiki/Encode/AAC/PNS_2.2.png
Reviewed-by: Claudio Freire <klaussfreire@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This patch fixes a pointer arithmetic bug in adjust_frame_information that resulted in heavily corrupted audio when using M/S encoding. Also, a backup copy of untransformed coefficients has to be kept around or attempts at re-processing the frame (which happens when hevavily overspending bits during transients) will result in re-encoding of the coefficients and subsequent corruption of the resulting stream.
A/B testing shows the bug as corrected, but still cannot prove that M/S coding is a win at least in numbers. Limited listening tests do show improvement on M/S encoded samples in lower bitrates, but they're hidden among the other artifacts that remain to be corrected in the encoder.
Some of the regressions flagged in the report do show poor stereo image (but not buggy), so M/S encoding is clearly not good enough yet to be defaulted to auto.
In numbers, Patched against Unpatched, stereo_mode auto:
Files: 114
Bitrates: 6
Tests: 683
Serious Regressions: 0 (0%)
Regressions: 0 (0%)
Improvements: 227 (33%)
Big improvements: 92 (13%)
Worst regression - mybloodrusts.wv - 256k
- StdDev: 28.61 pSNR: -0.43 maxdiff: 1372.00
Best improvement - 60.wv - 384k
- StdDev: -369.57 pSNR: 45.02 maxdiff: -13322.00
Average - StdDev: -80.56 pSNR: 2.49 maxdiff: -8858.00
Patched against Unpatched stereo_mode ms_off shows no difference.
Patched stereo_mode auto vs Unpatched stereo_mode ms_off shows a small average improvement, just not too significant:
Serious Regressions: 0 (0%)
Regressions: 10 (1%)
Improvements: 45 (6%)
Big improvements: 2 (0%)
Worst regression - Illinois.wv - 256k
- StdDev: 33.20 pSNR: -2.03 maxdiff: 477.00
Best improvement - song_of_circomstances.flac - 384k
- StdDev: -3.97 pSNR: 7.61 maxdiff: -826.00
Average - StdDev: -10.25 pSNR: 0.20 maxdiff: -281.00
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This fixes a case where multichannel bitrate isn't accurately
targetted by psy model alone, never achieving the target bitrate.
Signed-off-by: Martin Storsjö <martin@martin.st>
Fixes a case where multichannel bitrate isn't accurately
targetted by psy model alone, never achieving the target bitrate.
Now fixed.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Do not pointlessly call ff_alloc_packet multiple times,
and fix an infinite loop by clamping the maximum
number of bits to target in the algorithm that does
not use lambda.
Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de>
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Do not pointlessly call ff_alloc_packet2 multiple times,
and fix an infinite loop by clamping the maximum
number of bits to target in the algorithm that does
not use lambda.
Signed-off-by: Reimar Döffinger <Reimar.Doeffinger@gmx.de>
On cygwin, math.h needs to be included before float.h because of a bug
in the system headers. Including libavutil/libm.h first works around
this issue.
Longer discussion of the topic:
http://thread.gmane.org/gmane.comp.video.ffmpeg.devel/128582
Reduce scalefactors in non-zero bands that underflow by twice as much as those
in bands that just fail to hit psy targets.
Originally committed as revision 24482 to svn://svn.ffmpeg.org/ffmpeg/trunk
The minimal codebook to encode the band without clipping is used (as is done in the TLS).
Originally committed as revision 23318 to svn://svn.ffmpeg.org/ffmpeg/trunk
James Almer
Rostislav Pehlivanov
Claudio Freire
Peter Meerwald
Reimar Döffinger
Timothy Gu
Diego Biurrun
Michael Niedermayer
Nathan Caldwell
Mans Rullgard
Alex Converse
Reinhard Tartler
Young Han Lee
