TNS was computing filter coefficients incorrectly for short windows
due to a few coefficient addressing bugs. Fixing them fixes lots of
instability with transients (short windows).
With only 7 coefficients per short window at most the extra precision
makes a difference and seems to reduce crackling and stddev even
further.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
There are a couple of major changes here:
1. Start using TNS coefficient compression.
2. Start using 3 bits per coefficient maximum for short windows.
The bits we save from these 2 changes seem to make a nice impact on the
rest of the file/windows.
3. Remove special case gain checking for short windows.
4. Modify the coefficient loop to support up to 3 windows.
The additional restrictions on TNS were something that was no in the
specifications and furthermore restricting TNS to only low energy short
windows was done to compensate for bugs elsewhere in the code.
Overall, the improvements here reduce crackling artifacts heard in very
noisy tracks.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Changes:
- strongly prefer dual filters to a single filter
- less strict about using 2 filters w.r.t. energy
- scrap the usage of threshold and spread, useless
- use odd-shaped windows to set the filter direction
- use 4 bits instead of 3 bits for short windows
- simplify and reduce the main loop to a single level
- add stricter regulations for short windows
All of this now makes the TNS implementation operate
as good as it can and it definitely shows. The frequency
thresholds are now even better defined by looking at
the spectrals and the overall sound has been improved at
the price of just a few bits that are well worth it.
Too much effort and work has been spent on such a simple function.
It simply refuses to work as the specifications say, the
transformation is NOT lossless and creates some crackling and
distortions.
Therefore disable it by default and add a couple of warnings to
scare people away from touching it or wasting their time the
way I did.
The decoder does this so I guess we better do that as well.
There's barely any difference between the autoregressive and
the moving average filters looking at spectrals though.
This commit was made possible with the earlier commits since the
new quantization method basically means we're working always with
unsigned values. The specifications mention to use compression when
the first 2 bits are identical but they didn't mention if this should
happen before or after the conversion to signed values. Actually
they said nothing about conversion to signed values.
With this commit, coefficient compression usually always happens
which saves a lot of space, especially at extremely low bitrates
and doesn't change the quality at all.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This finally (and again) gets rid of basically everything the
specifications say about how TNS should be done. The main
problem used to be that a single filter was used for all
coefficients which despite being explicitly recommended by
the specifications usually sounds wrong, therefore it's
a corner case in the current TNS implementation.
This commit also changes the coefficient bit size, as apparently
it's better to use lower precision in case the windows are eight
short. This is apparently what fdk_aac uses, looking at the bit
stream and makes sense. Also the order when 8 SHORT windows happen
is important as 7 was too much and according to PSNR was worse
while 5 is just about correct.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit improves the TNS implementation to the point where it's
actually usable and very rarely results in nastyness (in all bitrates
except extremely low bitrates it's increasing the quality and prevents
some distortions from the coder being audiable).
Also adds a double filter support which is only used if the energy
difference between the top and bottom of the SFBs is above the
thresholds defined in the header file. Looking at the bitstream
that fdk_aac generates it sometimes used a double filter despite
the specs stating that a single filter should be enough for almost
all cases and purposes.
Unlike FAAC or fdk_aac we sometimes use a reverse filter in case
the energy difference isn't enought to use a double filter. This
actually works better.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit reworks the TNS implementation to a hybrid between what
the specifications say, what the decoder does and what's the best
thing to do.
The filter application function was copied from the decoder and
modified such that it applies the inverse AR filter to the
coefficients. The LPC coefficients themselves are fed into the
same quantization expression that the specifications say should
be used however further processing is not done, instead they're
converted to the form that the decoder expects them to be in
and are sent off to the compute_lpc_coeffs function exactly the
way the decoder does. This function does all conversions and will
return the exact coefficients that the decoder will generate, which
are then applied to the coefficients.
Having the exact same coefficients on both the encoder and decoder
is a must since otherwise the entire sfb's over which the filter
is applied will be attenuated.
Despite this major rework, TNS might not work fine on some audio
types at very low bitrates (e.g. sub 90kbps) as it can attenuate
some coefficients too much. Users are advised to experiment with
TNS at higher bitrates if they wish to use this tool or simply
wait for the implementation to be improved.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This was copied from the decoder, but is unneeded for the encoder.
tns_max_bands is unused and set to zero which zeroed out start, end
and size and thus no filter was actually applied.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Since the coefficients are stepped up to order + 1 it was possible
that it went over TNS_MAX_ORDER. Also just return in case the only
coefficient is less than the threshold.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The encoder-side filter isn't that important. The PSNR
shouldn't change so the FATE test should still be fine.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
The order should never go above TNS_MAX_ORDER (and thus cause
the context to be reinitialized) but this is just in case.
Also fix a comparison, since the coefficients are zero-indexed.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
It also made no sense to actually make the filter span the entire
window including the first band of the next window.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit abandons the way the specifications state to
quantize the coefficients, makes use of the new LPC float
functions and is much better.
The original way of converting non-normalized float samples
to int32_t which out LPC system expects was wrong and it was
wrong to assume the coefficients that are generated are also
valid. It was essentially a full garbage-in, garbage-out
system and it definitely shows when looking at spectrals
and listening. The high frequencies were very overattenuated.
The new LPC function performs the analysis directly.
The specifications state to quantize the coefficients into
four bit index values using an asin() function which of course
had to have ugly ternary operators because the function turns
negative if the coefficients are negative which when encoding
causes invalid bitstream to get generated.
This deviates from this by using the direct TNS tables, which
are fairly small since you only have 4 bits at most for index
values. The LPC values are directly quantized against the tables
and are then used to perform filtering after the requantization,
which simply fetches the array values.
The end result is that TNS works much better now and doesn't
attenuate anything but the actual signal, e.g. TNS removes
quantization errors and does it's job correctly now.
It might be enabled by default soon since it doesn't hurt and
helps reduce nastyness at low bitrates.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Thanks to @nevcairiel for pointing this one out.
Another thing which stopped msvc from compiling.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Hotfix to deal with msvc. Sane compilers lack POSIX ffs().
It only saves a single bit or so and isn't worth it that much.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
This commit implements temporal noise shaping support in the
encoder, along with an -aac_tns option to toggle it on or off
(off by default for now). TNS will increase audio quality
and reduce quantization noise by applying a multitap FIR filter
across allowed coefficients and transmit side information to the
decoder so it could create an inverse filter.
Users are encouraged to test the new functionality by enabling
-aac_tns 1 during encoding.
No major bugs are observable at this time so after a while if no
new problems appear and if the current implementation is deemed
of high enough quality and stability it will be enabled by default,
possibly at the same time the encoder has its experimental flag
removed and becomes the standard aac encoder in ffmpeg.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
