Hydrogenaudio Forums > Small benchmark with multichannel content

Full Version: Small benchmark with multichannel content

Hydrogenaudio Forums > Lossless Audio Compression > Lossless / Other Codecs

skamp

May 8 2005, 06:46

I've made a small benchmark of lossless codecs on 5.1 content instead of stereo. I tested FLAC 1.1.2, WavePack 4.2 and TTA 3.2, which are the only lossless codecs that I could use under linux and that support multichannel audio. The audio file is the song Roundabout by Yes, from the album "Fragile". I ripped it from the 5.1 DTS track of the video part of the audio DVD.

The results are surprising: TTA seems to be much less efficient than with stereo audio, FLAC -q 2 is *faster* than -q 0 and -q 1, and compresses *better* than -q 0-6.
Here is a modest chart (I'm allergic to office suites), and here is the values I gathered, in CSV format:

CODE

codec,bytes,compression,encoding time,encoding speed,decoding time,decoding speed
wav,295768064,100,513.482,1,513.482,1
flac -8,142495744,48.178%,110.944,4.628,14.036,36.583
flac -7,143101952,48.383%,92.784,5.534,11.868,43.266
flac -6,157319168,53.190%,30.281,16.957,13.648,37.623
flac -5,158875648,53.716%,29.002,17.705,12.712,40.393
flac -4,158875648,53.716%,28.541,17.991,13.638,37.650
flac -3,159535104,53.939%,27.801,18.469,12.802,40.109
flac -2,155369472,52.530%,23.062,22.265,11.347,45.252
flac -1,163622912,55.321%,24.584,20.886,11.610,44.227
flac -0,163622912,55.321%,23.227,22.107,11.705,43.868
tta,194715648,65.833%,25.870,19.848,22.207,23.122
wavpack -hx2,128733184,43.525%,335.459,1.530,50.823,10.103
wavpack -h,132235264,44.709%,57.189,8.978,56.282,9.123
wavpack -x4,131035136,44.303%,585.412,.877,24.622,20.854
wavpack -x2,132345856,44.746%,138.210,3.715,27.779,18.484
wavpack,134819840,45.582%,35.536,14.449,25.849,19.864
wavpack -fx6,134098944,45.339%,316.097,1.624,19.707,26.055
wavpack -fx4,134250496,45.390%,186.806,2.748,20.649,24.867
wavpack -fx2,136560640,46.171%,78.838,6.513,19.346,26.542
wavpack -f,139022336,47.003%,26.445,19.416,19.167,26.789

This benchmark isn't meant to be comprehensive of course. Feel free to take the time to make a more complete one, it would be very welcome.

guruboolez

May 8 2005, 07:52

Very interesting - and also comprehensive in my opinion! It would be interesting to confirm (or not) the poor results of TTA encoder with multichannel. I have currently a DTS album on my computer (AERO, from Jean-Michel Jarre), and if I find space enough on my HDD, I'll try to observe the reaction of the different encoders.

Just one question: are you running with Windows XP? It would be nice to see the performance of WMA9 lossless compared to other format.

jaybeee

May 8 2005, 12:47

QUOTE(guruboolez @ May 8 2005, 02:52 PM)

Just one question: are you running with Windows XP? It would be nice to see the performance of WMA9 lossless compared to other format.

QUOTE(skamp @ May 8 2005, 01:46 PM)

I tested FLAC 1.1.2, WavePack 4.2 and TTA 3.2, which are the only lossless codecs that I could use under linux and that support multichannel audio.

Guess not

guruboolez

May 9 2005, 03:18

Anyway, I've encoded the J.M. Jarre' Aero album. DTS format@48000Hz@6channels.

Results:

CODE

FORMAT FILESIZE BITRATE ALBUM LENGTH

dts 830.814 kb 1536 kbps¹ 1:13:51
flac5 1.176.862 kb 2176 kbps¹ 1:13:51
tta 1.367.129 kb 4294963595¹ 1:13:51
wv4 -f 1.431.884 kb 3146 kbps¹ 1:02:08 (!)

¹ foobar2000 value [TTA value is incoherent]

Strange things happened.

• bitrate reported by foobar2000 is often incoherent (if flac is 2176 kbps for 1176 MB, wv can't be at 3146 kbps for 1431 Mb ; TTA bitrate is nonsense).

• the wavpack encoding is 10 minutes shorter (I haven't checked if music is really missing - I was in hurry this morning)

• I've transcoded it with foobar2000, and a bit-to-bit comparison told me that files are not bit-to-bit identical.

It is interesting to note flac good ratio ; TTA is clearly behind, and wavpack even farer.

rutra80

May 9 2005, 04:52

QUOTE(guruboolez @ May 9 2005, 11:18 AM)

• I've transcoded it with foobar2000, and a bit-to-bit comparison told me that files are not bit-to-bit identical.

Maybe because of dithering (DTS is lossy)?

The Link

May 9 2005, 05:33

@guruboolez: Did you use the "-i" switch for wavpack with foo_clienc? Without it, encoding even failed here but when using it, the reported length and bitrate seem to be OK (though the resulting file isn't bit identical, too).

edit: with the -i switch you don't need the "encoder requires accurate length" option to be ticked anymore.

guruboolez

May 9 2005, 05:48

rutra80>
I'll see when I return at home. I didn't perform a complete bit-to-bit comparison this morning ; I've just saw that difference appareared during the first seconds, and cancelled the (slow) operation. I've maybe choose bad options in diskwriter (DSP was off, I've verifed).

The Link>
I'm 99% sure that -i command wasn't active.

pest

May 9 2005, 06:00

using a lossy compressed source isn't a good opinion.
the noise introduced degrades results and the comparision is
a bit senseless because the output is larger

the only way to compare multichannel abilities of the different
encoders is to use a orginal multichannel-pcm source

i hope that doesn't sound too harsh

greets
pest

guruboolez

May 9 2005, 06:08

pest> you're right. But the purpose of my test was to see if skamp's results -especially for TTA- could be reproduced with another multichannel album. In other word, checking the efficiency of multichannel encoding with various format. With only two albums, we have totally different results:

• skamp: WavPack > flac > TTA
• guruboolez: flac > TTA > WavPack

Consequently, we can't say anything about efficiency with multichannel content for these three formats.

pest

May 9 2005, 06:19

one explanation for the different results
this is only speculation

flac has a better way to detect correct coupling of channels and doesn't
assumes a fixed order as wavpack does (correct me if i'm wrong)
tta's channel coupling techniques are low as i checked the source a while back

wavpack and tta rely on adaptive prediction but flac has a more robust
forward prediction sheme which could compensate for the noise

anyway thanks for the work, because large multichannel-files seem to show
some strange behavior in current implementions

This is a "lo-fi" version of our main content. To view the full version with more information, formatting and images, please click here.