[Ffmpeg-devel] FLAC encoder

[Justin Ruggles]

Hi,

My latest version of Flake is almost ready to attempt an FFmpeg
integration patch.  Development is going slower than expected, but I
only have a couple things left on my list now.  I am still working on
blocksize switching, and I also want to add MD5 checksum support
sometime soon.

I have made a small webpage for updates and test results.

http://home.earthlink.net/~jruggle/flake/

A quick run-down of some of the changes:

New preset compression levels:
Now 0-8 more closely mirror the reference encoder.  Josh Coalson made a
good point on Hydrogenaudio that my encoder was relying too heavily on
higher block sizes and prediction orders, which increase decoding time &
decoding buffer requirements.  Making this change also allows an
apples-to-apples comparison.  I have put in levels 9-15 for higher block
sizes, orders, and slower-but-better compression modes.

New stereo decorrelation:
Added a search method which is slower, but much more accurate.  This is
used in compression levels 6 and above.  Estimation is still used for 5
and below.  Encoding without decorrelation is also a commandline option.

Better prediction order selection:
The old method picked the maximum order about 95% of the time, which was
not very good.  I added a new estimation method, which is okay but not
spectacular.  The best addition was the quad method.  It selects a good
estimate by doing bit counts for the maximum order, 3/4 max, 1/2 max,
and 1/4 max.

More commandline options:
usage: flake [options] <input.wav> <output.flac>
options:
       [-0 ... -15] Compression level (default: 5)
                        0 = -b 512   -l 0  -o 0 -s 1
                        1 = -b 512   -l 4  -o 0 -s 1
                        2 = -b 1024  -l 4  -o 1 -s 1
                        3 = -b 2048  -l 5  -o 1 -s 1
                        4 = -b 4096  -l 6  -o 1 -s 1
                        5 = -b 4608  -l 8  -o 1 -s 1
                        6 = -b 4608  -l 8  -o 2 -s 2
                        7 = -b 4608  -l 8  -o 3 -s 2
                        8 = -b 4608  -l 12 -o 3 -s 2
                        9 = -b 4608  -l 12 -o 4 -s 2
                       10 = -b 4608  -l 32 -o 3 -s 2
                       11 = -b 4608  -l 32 -o 4 -s 2
                       12 = -b 8192  -l 32 -o 3 -s 2
                       13 = -b 8192  -l 32 -o 4 -s 2
                       14 = -b 16384 -l 32 -o 3 -s 2
                       15 = -b 16384 -l 32 -o 4 -s 2
       [-b #]       Block size [16 - 65535] (default: 4608)
       [-l #]       Maximum prediction order [0 - 32] (default: 8)
       [-o #]       Prediction order selection method
                        0 = maximum
                        1 = estimate (default)
                        2 = dual
                        3 = quad
                        4 = search
       [-s #]       Stereo decorrelation method
                        0 = independent L+R channels
                        1 = estimate (default)
                        2 = search

I am now considering removing the levels with 16384 block size because
they don't seem to be performing as well as before.  I think my encoder
just got better at other block sizes.  Also, the 4-level search doesn't
seem to have a huge advantage when used with max order 32...it's only a
little faster, but usually has too big of a compression loss.  I want to
add an 8-level order search as another option for use with high max
orders.  It would give speed similar to level 9, but with better
compression.

Thanks,
-Justin