Sadly the ISO frame size limit on LAME 3.98.2 causes regression with 320kbps CBR.
...LAME 3.98 sounds the worst, due to its harsher pre-echo artifacts. ...
Do you mind trying CBR 256 and ABR 270 (or similar)?
As bit reservoir is used with these settings there may be a chance that audio data amount at the critical spots is higher than when using CBR 320.
I've tried out 256 and 270 ABR (I got a 285 encode) with a more critical sample:
LAME 3.98.2 256 vs 320
foo_abx 1.3.4 report
foobar2000 v1.0
2010/01/23 21:31:45
File A: C:\Temp\musique_non_stop 256 CBR.mp3
File B: C:\Temp\musique_non_stop 320 CBR.mp3
21:31:45 : Test started.
21:32:26 : 00/01 100.0%
21:32:52 : 01/02 75.0%
21:33:13 : 02/03 50.0%
21:33:52 : 03/04 31.3%
21:34:21 : 04/05 18.8%
21:34:44 : 04/06 34.4%
21:35:00 : 05/07 22.7%
21:35:25 : 06/08 14.5%
21:35:47 : 07/09 9.0%
21:36:02 : 08/10 5.5%
21:36:24 : 09/11 3.3%
21:36:47 : 10/12 1.9%
21:36:50 : Test finished.
----------
Total: 10/12 (1.9%)
The 256 encode sounds the worse, but the 320 version sounds only a little bit better.
LAME 3.98.2 270 ABR vs 320 CBR
foo_abx 1.3.4 report
foobar2000 v1.0
2010/01/23 21:38:52
File A: C:\Temp\musique_non_stop 270 ABR.mp3
File B: C:\Temp\musique_non_stop 320 CBR.mp3
21:38:52 : Test started.
21:39:33 : 00/01 100.0%
21:40:03 : 01/02 75.0%
21:40:14 : 02/03 50.0%
21:40:23 : 03/04 31.3%
21:40:41 : 04/05 18.8%
21:40:55 : 05/06 10.9%
21:41:08 : 06/07 6.3%
21:41:34 : 07/08 3.5%
21:41:44 : 08/09 2.0%
21:41:58 : 09/10 1.1%
21:42:23 : 10/11 0.6%
21:42:31 : 11/12 0.3%
21:42:39 : 12/13 0.2%
21:42:49 : 13/14 0.1%
21:42:58 : 14/15 0.0%
21:43:01 : Test finished.
----------
Total: 14/15 (0.0%)
The 270 ABR versions is at 285kbps, sadly it sounds worse then 320kbps.
LAME 3.98.2 256 CBR vs 270 ABR
foo_abx 1.3.4 report
foobar2000 v1.0
2010/01/23 21:53:36
File A: C:\Temp\musique_non_stop 256 CBR.mp3
File B: C:\Temp\musique_non_stop 270 ABR.mp3
21:53:36 : Test started.
21:53:53 : 01/01 50.0%
21:53:59 : 02/02 25.0%
21:54:05 : 03/03 12.5%
21:54:14 : 04/04 6.3%
21:54:21 : 05/05 3.1%
21:54:29 : 06/06 1.6%
21:54:37 : 07/07 0.8%
21:54:50 : 08/08 0.4%
21:54:59 : 09/09 0.2%
21:55:12 : 10/10 0.1%
21:55:20 : 11/11 0.0%
21:55:29 : 12/12 0.0%
21:55:30 : Test finished.
----------
Total: 12/12 (0.0%)
The 256 encode has horrid precho at 0:04, while the 270 version is more clearer but still very bad.