AAC - 96 kbps, Intensity Stereo Test Options

[Ivan Dimkovic]

For the upcoming brand new AAC encoder from Nero we just implemented Intensity Stereo, too

This was the very old discussion in AAC-Tech forum whether IS should be used or not - and, indeed, how much it degrades sound image and how much it brings in improved overal distortion-level...

So, here is the first test with Nero's new encoder which is currently in development (signalled as 4.9.9.9 this time

Files are encoded in 96 kbps CBR - to maximize the "pressure" on the codec psymodel/bit-rate-allocator - and following files are covered:

castanets -> pre echo clip
dr4 -> trumpet with strong harmonic components
es02 -> German speech
French_ad -> French Ad, also speech but with music in background
Layla -> strong applause in both channels that makes lossy stereo very tricky
Si02 -> pre echo clip
Sm01 -> highly tonal clip
Velvet -> speciality for stereo coding almost completely uncorrelated channels
Waiting -> we all know what is it...
youcantdothat -> also strong channel separation

I chose 96 kbps to measure how much IS degrades the quality, at bit rates of 64 kbps it is clear that IS definitely outweights any of its disadvantages... but as we go up, it is interesting to check.

Also I selected very worst test items, just to make it even harder for the codec

Note - IS implementation in this version is still not final, and it will for sure be improved even more before the launch.

Attached are packed IS and non-IS files (1.65 mb each)... original flacs coming soon

Edit - added files with better IS perceptual control

This post has been edited by Ivan Dimkovic: Dec 25 2005, 19:37

Attached File(s)

 aac_96kbps_cbr_no_is.rar ( 1.66MB ) Number of downloads: 391
 aac_96kbps_cbr_with_is.rar ( 1.65MB ) Number of downloads: 380
 aac_96kbps_cbr_with_is2.rar ( 1.65MB ) Number of downloads: 399

 

[Ivan Dimkovic]

And... the flacs

Attached File(s)

 castanets.flac ( 621.81K ) Number of downloads: 458
 dr4.flac ( 1.16MB ) Number of downloads: 338
 es02.flac ( 482.46K ) Number of downloads: 349
 French_Ad.flac ( 912.28K ) Number of downloads: 369
 Layla.flac ( 2.2MB ) Number of downloads: 394
 sm01.flac ( 1MB ) Number of downloads: 376
 VELVET.flac ( 1.36MB ) Number of downloads: 379
 Waiting.flac ( 2.29MB ) Number of downloads: 380
 youcantdothat.flac ( 3.43MB ) Number of downloads: 422
 si02.flac ( 625.48K ) Number of downloads: 538

 

[Ivan Dimkovic]

Another note - most likely, IS won't be used at this bit rate, or it will be used in much reduced way than it is in this test - purpose of the demo is to check amount of stereo reduction (if any) and some other artifacts it might introduce.

Few of those are already known and we are working on them

So, at lower bitates (e.g. 48-80 kbps) we might improve listening experience over common LC-AAC implementations even more for those without ability to use HE-AAC

This post has been edited by Ivan Dimkovic: Dec 25 2005, 02:46

[guruboolez]

Short ABX tests before jumping in a warm bed.
Keep in mind that I haven't PCM files as reference; I only compared both encodings and tried to find encoding artifacts.

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/25 03:08:59

File A: C:\HA - noël 2005\NERO Intensity Stereo\french_Ad.mp4
File B: C:\HA - noël 2005\NERO Intensity Stereo\french_Ad_IS.mp4

03:08:59 : Test started.
03:09:40 : 01/01 50.0%
03:09:47 : 02/02 25.0%
03:09:54 : 03/03 12.5%
03:09:59 : 04/04 6.3%
03:10:05 : 05/05 3.1%
03:10:11 : 06/06 1.6%
03:10:16 : 07/07 0.8%
03:10:25 : 08/08 0.4%
03:10:27 : Test finished.

----------
Total: 8/8 (0.4%)

The IS encoding sounds noisier on the beginning (little pshhh audible on both encodings but slightly higher with IS). Difference is rather small.

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/25 03:13:19

File A: C:\HA - noël 2005\NERO Intensity Stereo\dr4.mp4
File B: C:\HA - noël 2005\NERO Intensity Stereo\dr4_IS.mp4

03:13:19 : Test started.
03:13:44 : 01/01 50.0%
03:13:52 : 02/02 25.0%
03:14:02 : 03/03 12.5%
03:14:08 : 04/04 6.3%
03:14:18 : 05/05 3.1%
03:14:37 : 06/06 1.6%
03:14:44 : 07/07 0.8%
03:15:04 : 07/08 3.5%
03:15:10 : 08/09 2.0%
03:15:15 : 09/10 1.1%
03:15:17 : Test finished.

----------
Total: 9/10 (1.1%)

[11.00 - 13.00] The loudest trumpet note is distorted (it sounds like ringing on a very short moment) with IS encoding. The other one is less disturbing.
Additional note: the distortion is also audible on the beginning (two first seconds). This time both encodings are concerned, but IS is really more unpleasant to my ears. It's not really ringing, but noise or fat sounding.
Additional note No.2: sample sm01 suffers from ringing (more with IS), but I haven't ABXed it yet.

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 12
2005/12/25 03:15:48

File A: C:\HA - noël 2005\NERO Intensity Stereo\castanets.mp4
File B: C:\HA - noël 2005\NERO Intensity Stereo\castanets_IS.mp4

03:15:48 : Test started.
03:16:12 : 01/01 50.0%
03:16:18 : 02/02 25.0%
03:16:24 : 03/03 12.5%
03:16:38 : 04/04 6.3%
03:16:43 : 04/05 18.8%
03:16:50 : 05/06 10.9%
03:17:00 : 05/07 22.7%
03:17:05 : 06/08 14.5%
03:17:10 : 07/09 9.0%
03:17:23 : 07/10 17.2%
03:17:34 : 08/11 11.3%
03:17:46 : 08/12 19.4%
03:17:54 : 09/13 13.3%
03:18:30 : 09/14 21.2%
03:18:35 : 10/15 15.1%
03:18:38 : 10/16 22.7%
03:18:43 : 10/17 31.5%
03:18:45 : Test finished.

----------
Total: 10/17 (31.5%)

I heard more smearing for IS during the introduction (guitar + ta-ta-ta-ta) but I decided to ABX something which isn't located on the very beginning. I failed :/

This is all for this night. Now, bed. Tomorrow, some other files
P.S. The german speech sample sounds really bad, with aggressive whispering I can sometimes heard on TV. Are your encodings transcoded from a lossy source?

P.S.2: hardware setting = AC97 component + Philips HP910 headphone

This post has been edited by guruboolez: Dec 25 2005, 03:33

[Ivan Dimkovic]

Thanks Guru for the quick test! Also warm bed is waiting me, too

Regarding ES02... please download the .flac file - it is not transcoded or processed in any way, and encoded files are very close to it.

[guruboolez]

You're right: the reference disturbs me as well (only the 00.600-04.600 range - the rest sounds OK).

[Ivan Dimkovic]

Some more perceptual optimizations are done - I'll try to post new IS clips at the same bitrate soon.

[Ivan Dimkovic]

Ok, here are the files with refined IS algo (more perceptual control)

Attached File(s)

 aac_96kbps_cbr_with_is2.rar ( 1.65MB ) Number of downloads: 287

 

[IgorC]

It's test really not for 10$ edifier 6631 and generic audio card

CODE

ABC/HR Version 1.1 beta 2, 18 June 2004
Testname: IS2 and NO_IS

1L = C:\music\waiting abx\1 Waiting is2.wav
2L = C:\music\waiting abx\2 Waiting no is.wav

 0 of   1, p = 1.000
 0 of   2, p = 1.000
 1 of   3, p = 0.875
 2 of   4, p = 0.688
 3 of   5, p = 0.500
 4 of   6, p = 0.344
 5 of   7, p = 0.227
 5 of   8, p = 0.363
FINISHED


---------------------------------------
General Comments:

---------------------------------------
ABX Results:
C:\music\waiting abx\1 Waiting is2.wav vs C:\music\waiting abx\2 Waiting no is.wav
   5 out of 8, pval = 0.363

This post has been edited by IgorC: Dec 25 2005, 22:43

[IgorC]

So, at lower bitates (e.g. 48-80 kbps) we might improve listening experience over common LC-AAC implementations even more for those without ability to use HE-AAC 

CT supports HE-AAC up to 128 kbit/s. After some abx test ( Winamp 5.12 CT encoder vs Nero 4.9.9.9 for the samples of this topic ) I found that CT encoder can be on par with LC-AAC at 96 kbit/s or even beter on some samples like Waiting and Layla

Indeed there should be a test to determine if statement that SBR is usefull only up to 80 kbit/s is still truely.

This post has been edited by IgorC: Dec 26 2005, 05:35

[Ivan Dimkovic]

For 96kbps there is also choice of single-rate SBR (downsampled SBR) - where AAC codec would operate at 44.1 kHz - thus eliminating SBR pre-echo issues a lot.

Our new codec also supports that, and it would be very interesting to check if this mode outperforms LC-AAC and usual dual-rate HE-AAC at this bit rate.

This post has been edited by Ivan Dimkovic: Dec 26 2005, 11:30

[Sebastian Mares]

Some quick tests:

No IS vs. Original:

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 13
2005/12/26 11:53:21

File A: D:\Neuer Ordner\aac_96kbps_cbr_no_is\Layla.mp4
File B: D:\Neuer Ordner\Layla.flac

11:53:21 : Test started.
11:53:33 : 01/01 50.0%
11:53:39 : 02/02 25.0%
11:54:44 : 03/03 12.5%
11:54:48 : 04/04 6.3%
11:54:57 : 05/05 3.111:55:00 : 06/06 1.6%
11:55:11 : 07/07 0.8%
11:55:19 : 08/08 0.4%
11:56:01 : Test finished.

----------
Total: 8/8 (0.4%)

IS vs. Original:

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 13
2005/12/26 11:59:29

File A: D:\Neuer Ordner\Layla.flac
File B: D:\Neuer Ordner\aac_96kbps_cbr_with_is\Layla.mp4

11:59:29 : Test started.
11:59:37 : 01/01 50.0%
11:59:44 : 02/02 25.0%
11:59:47 : 03/03 12.5%
11:59:51 : 04/04 6.3%
11:59:54 : 05/05 3.1%
12:00:03 : 06/06 1.6%
12:00:06 : 07/07 0.8%
12:00:09 : 08/08 0.4%
12:00:10 : Test finished.

----------
Total: 8/8 (0.4%)

IS2 vs. Original:

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 13
2005/12/26 12:00:42

File A: D:\Neuer Ordner\Layla.flac
File B: D:\Neuer Ordner\aac_96kbps_cbr_with_is2\Layla.mp4

12:00:42 : Test started.
12:00:51 : 01/01 50.0%
12:00:54 : 02/02 25.0%
12:00:57 : 03/03 12.5%
12:00:59 : 04/04 6.3%
12:01:01 : 05/05 3.1%
12:01:03 : 06/06 1.6%
12:01:04 : 07/07 0.8%
12:01:06 : 07/08 3.5%
12:01:12 : Test finished.

----------
Total: 7/8 (3.5%)

No IS vs. IS:

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 13
2005/12/26 12:02:00

File A: D:\Neuer Ordner\aac_96kbps_cbr_no_is\Layla.mp4
File B: D:\Neuer Ordner\aac_96kbps_cbr_with_is\Layla.mp4

12:02:00 : Test started.
12:02:14 : 01/01 50.0%
12:03:00 : 02/02 25.0%
12:03:18 : 03/03 12.5%
12:03:22 : 04/04 6.3%
12:03:33 : 05/05 3.1%
12:04:04 : 06/06 1.6%
12:04:10 : 07/07 0.8%
12:04:13 : 07/08 3.5%
12:04:15 : Test finished.

----------
Total: 7/8 (3.5%)

No IS vs. IS2:

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 13
2005/12/26 12:06:10

File A: D:\Neuer Ordner\aac_96kbps_cbr_no_is\Layla.mp4
File B: D:\Neuer Ordner\aac_96kbps_cbr_with_is2\Layla.mp4

12:06:10 : Test started.
12:06:31 : 01/01 50.0%
12:06:34 : 01/02 75.0%
12:06:38 : 02/03 50.0%
12:06:40 : 03/04 31.3%
12:06:54 : 03/05 50.0%
12:07:09 : 04/06 34.4%
12:07:13 : 04/07 50.0%
12:07:16 : 05/08 36.3%
12:07:18 : Test finished.

----------
Total: 5/8 (36.3%)

IS vs. IS2:

CODE

foo_abx 1.3 report
foobar2000 v0.9 beta 13
2005/12/26 12:07:47

File A: D:\Neuer Ordner\aac_96kbps_cbr_with_is\Layla.mp4
File B: D:\Neuer Ordner\aac_96kbps_cbr_with_is2\Layla.mp4

12:07:47 : Test started.
12:08:33 : 00/01 100.0%
12:08:36 : 00/02 100.0%
12:08:41 : 00/03 100.0%
12:08:44 : 01/04 93.8%
12:08:47 : 01/05 96.9%
12:08:51 : 02/06 89.1%
12:08:53 : 03/07 77.3%
12:08:56 : 04/08 63.7%
12:08:58 : Test finished.

----------
Total: 4/8 (63.7%)

Conclusion: all are more or less distinguishable from the original. When it comes to IS vs. no IS, no IS sounds better to my ears. For IS vs. IS2, no idea. I think IS2 was a bit better, but as you can see from the ABX logs, I didn't manage to get an acceptable result.

I also noticed that IS has an artifact during the first few seconds of applause. It produces something like a "Krr".

This post has been edited by Sebastian Mares: Dec 26 2005, 12:21

[Ivan Dimkovic]

Thanks IgorC, Sebastian,

We will continue to improve IS (some more things are left to be done) - but I also think IS should be used at bit rates lower than 96 kbps - perhaps 80 and definitely 64 kbps.

Interesting is that I definitely found out that IS in AAC is much better than IS in MP3 - because AAC has TNS tool, which should shape the noise to avoid IS articacts of bad noise positioning in the R channel (thus generating failed stereo image)

[freelancer]

Glad to hear Nero is doing something new to AAC, I've a question that is the AAC+IS require the IS support from AAC decoder just like SBR and PS does?
My new mobile phone seems to only support LC-AAC, so I can't get any benifit from HE-AAC, if LC+IS doesn't require decoder's special support, that could be cool to me.

[Garf]

IS is a part of the original, very first, MPEG 2 LC-AAC spec, so *all* AAC decoders have support for it.

[freelancer]

IS is a part of the original, very first, MPEG 2 LC-AAC spec, so *all* AAC decoders have support for it.

Thanks Garf, glad to hear that. But why IS is already supported by decoders but no encoders support IS till now?(except the coming Nero encoder )
And when can we expect to have a public test version of Nero AAC encoder that supports IS? And did you do any internal comparison about LC+IS and HE-AAC at 64/80 Kbps? I'm interested at which one could win
Or it is possible to use HE+IS+PS in the future?

This post has been edited by freelancer: Jan 7 2006, 13:05

[Garf]

IS is a part of the original, very first, MPEG 2 LC-AAC spec, so *all* AAC decoders have support for it.

Thanks Garf, glad to hear that. But why IS is already supported by decoders but no encoders support IS till now?(except the coming Nero encoder )

It appears to be quite tricky to use correctly.

And when can we expect to have a public test version of Nero AAC encoder that supports IS? And did you do any internal comparison about LC+IS and HE-AAC at 64/80 Kbps? I'm interested at which one could win
Or it is possible to use HE+IS+PS in the future?

We are still looking into this. For close calls we may do further public tests here.

[gameplaya15143]

For 96kbps there is also choice of single-rate SBR (downsampled SBR) - where AAC codec would operate at 44.1 kHz - thus eliminating SBR pre-echo issues a lot.

Our new codec also supports that, and it would be very interesting to check if this mode outperforms LC-AAC and usual dual-rate HE-AAC at this bit rate.

so does single rate/downsampled sbr mean that the lc-aac part would be 44.1khz and the sbr part would also be 44.1khz?

if so, i think it would outperform lc-aac and dual rate he-aac

[Garf]

so does single rate/downsampled sbr mean that the lc-aac part would be 44.1khz and the sbr part would also be 44.1khz?

SBR runs at 88.2kHz but gets downsampled (hence the name).

if so, i think it would outperform lc-aac and dual rate he-aac

I don't think anything, I am SURE it needs to be TESTED.

This post has been edited by Garf: Mar 5 2006, 10:39

[gameplaya15143]

SBR runs at 88.2kHz but gets downsampled (hence the name).

I don't think anything, I am SURE it needs to be TESTED.

thanks, I get it now

sure it would need testing... but if the lc-aac part isn't forced down to half the samplerate (22.05khz) then sbr wont have as much bandwidth to cover (instead of the entire top half, maybe just the top fourth) my speculation here is based on plusV fullrate (44.1khz mp3 + 44.1khz plusV)... we shall see how it goes when it gets tested though