[Lame-dev] substep testing

[Gabriel B. <bou...@mp...>]

I have some test results related to substep5.
I was experimenting in order to find a replacement for preset standard. I
proposed several configurations to some HA people (because I do not have
exceptionnal hearing or sound setup).
The good point is that they liked standard3, which I will probably use as
standard replacement. That is a good thing because it seems to improve
quality most of the times over old standard, and is not using "internal
cooking".

Based on this one, I submitted standard4 for testing. Std4 is std3 +
substep5 + lower ath.
I lowered the ath in order to compensate for the bitrate drop (hoping it
would increase quality).

They found a lot of tracks where this configuration sounded bad. It seems to
introduce a kind of flanging.
" Standard4 is more than a codec : it's an artist, who likes vibrato "

I looked at some tracks they pointed out (an example is the french horn from
here : http://www.pcabx.com/product/mds-jb920/index.htm).
On this track, the bitrate is reduced a lot when using substep5. Lowering
the ath does not help, because even without any ath, the bitrate is still
quite lower.

I am a little worried about this, as a potential explanation could be the
following one:
Without substep, we are using more bits for sfb 0-20 than necessary
(according to the psymodel) because of the sfb21 problem. With substep5,
sfb0-20 are using less bits, and are more representative of the psymodel
requirements. The bad results could mean that on high vbr bitrates we only
have good quality because we are using more bits than necessary (according
to psymodel) on sfb0-20.

However substep1/5 might be usefull, even in the current implementation, for
abr/cbr (I did not tested it so much)



----
Gabriel Bouvigne
www.mp3-tech.org
personal page: http://gabriel.mp3-tech.org