There is a strange behaviour that is only obvious at 280+ k bitrate. Below is a comparison of v3.98 and older encoders - target bitrate (produced bitrate). Over 20 samples were tested. At 288k the psymodel uses max tunings as far as I know so a *slight* rise is normal. V3.98 bitrate looks very strange.


Lame 3.98: target 280 (291), 288 (316)
Lame 3.97: target 280 (269), 288 (282)
Lame 3.90: target 288 (289)

perhaps things are messed up due to the mix of vbr-old and vbr-new not fully switched?

i could be lying, tho.


later

This behaviour is because of the restricted bit reservoir usage. You can see a similar effect with 3.97 if you add --strictly-enforce-ISO

robert, that is what i initially posted, but i did some encodings and the behavoir was much different.

--strictly-enforce-iso adds more bloat to both 3.97 AND 3.98b4 and doesn't address the OP


later

You are right, the switch --strictly-enforce-ISO is a harder constrain than what we had to add to work around the FhG decoder bug. But the effect is similar.

v3.98

--preset 270 (274)
--preset 265 (266)
-V0 -Y (218) - 225 is probably average without sbf21 issue.

v3.97

--preset 270 (262)

I guess this is more inline with the expected bitrate. I am simply testing a 'braindead' setting that gives more headroom over -V0 yet much smaller than cbr320. Bitrate should be 260~280k range. Goal is to have the advantages of cbr320. Transcoding, DSP application like vocal cut , some problem samples are still issues for -V2 / V0. In my experience -b320 is usually excellent in these cases and so far 265k abr is a good candidate . I'll post more of my findings soon.

Quite interesting cause this is pretty much what I did when I had my archive encoded with 3.90.3. I used ABR 270 then and didn't encounter a disadvantage against CBR 320.