[FFmpeg-devel] [FFmpeg-devel-irc] IRC log for 2010-03-27

[Alexander Strange]

On Mar 30, 2010, at 8:20 PM, Frank Barchard wrote:

> On Tue, Mar 30, 2010 at 9:23 AM, Michael Niedermayer <michaelni at gmx.at>wrote:
> 
>> 
>> if wma sounds better than aac then there must be a pretty nasty bug in the
>> aac encoder
>> 
>> 
> Instead of 'sounds better', is there a good metric?

Audio encoding quality can't be tested with PSNR, and usually isn't tested with metrics at all.
LAME releases were tested by hydrogenaudio posters using ABX testing, though I don't know how they did development. I guess Gabriel would know.

See:
http://ff123.net/abchr/abchr.html
http://wiki.hydrogenaudio.org/index.php?title=Hydrogenaudio_Listening_Tests
http://www.rarewares.org/rja/ListeningTest.pdf

ff123's ABC/HR tool is written in Java, so people around here might not want to run it, even if they do want to train themselves into hearing audio artifacts.
I do seem to have an 'abxtest' program which will do ABX testing, looks like it came with madplay.

There is a metric for it called PEAQ but I don't think there are any good implementations of it available.


> Running tiny_psnr on some audio tracks, the numbers do look suspicious.
> mp3 stddev: 3965.61 PSNR: 32.59 bytes: 84256244/ 84266496 (lame/ffmpeg).
> size 11 MB
> ogg  stddev:  296.70 PSNR: 46.88 bytes: 84256244/ 84260096 (oggenc/ffmpeg).
> size 6 MB
> aac stddev: 4015.37 PSNR: 31.92 bytes: 84256244/ 84262912 (faac/ffmpeg).
> size 7.2 MB
> wma1 stddev: 3877.18 PSNR: 34.10 bytes: 84256244/ 84262912 (ffmpeg/ffmpeg).
> size 4 MB.
> he-aac stddev:  288.77 PSNR: 47.12 bytes: 84256244/ 84256240 (nero/nero).
> size 10 MB
> But PSNR doesn't account for phase shift, so it may be a bad metric for
> audio.
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at mplayerhq.hu
> https://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-devel