As many people have written, FhG is unbeatable at 128kbs. I think it is $200 for linux, cheaper for windows.

In my tests, Xing($20?) and LAME are pretty close and both give
acceptable results at 128kbs. I have samples where LAME does better
than Xing, and vice versa. (but I am probably biased).

BladeEnc produces output identical to the ISO dist10.tar code. Thus
it has not yet fixed many serious bugs in the psycho acoustics and bit
allocation routines. You can see a list of these bugs on the LAME web
page. One example: the pre-echo detection turns on the window
switching exactly one frame too late. Thus the pre-echo is completely
missed and the window switching causes more harm than good.


Since you are going to invest a lot of time into encoding, I think it
would be worth it to downloading the free encoders, and the demo
versions of the commercial encoders and do your own tests with a good
pair of headphones. Test the music you like to listen too, and listen
to short passages dozens of times. Some encoder artifacts you wont
notice at first, but once you "learn" to hear them, you'll really notice them. The LAME homepage has many test cases with notes
on what types of flaws to listen for and what techniques we have coded
to improve the quality.

Mark