After having switched from --r3mix to "--alt-preset standard -Y",
and being initially pleased with the comparable bitrate (and even
posting on HA as such), now I am starting to feel the loss in HF
detail from -Y is definitely noticible and even annoying at times
(esp. on acoustic/vocal music).

So, I did some more searching on HA and groups.google.com,
and came across some postings from others with similar findings
(-Y just loses too much HF detail) ... one person suggested using
"--alt-preset standard -V3" as a compromise between straight
APS and APS-Y, and after several test encodes I find this setting
to be quite good ... bitrates are slightly higher than APS-Y, but
still almost always <200kbps which is acceptable to me (APS at
220+kbps still seems bloated, due to well-known sfb21 issue).

I know Dibrom's APS was tested as-is, with VBR quality 2 (-V2)
but what is the real effect [on quality] of adding -V3 to APS?
Please don't just reply with "that is not tested, don't use it"
I would like to know from some people involved in actual coding
or maybe who have actually tested APS -V3 (or APS -V4, etc) if
this is a better/worse (or just different, but how) from APS-Y.

Thanks in advance for your insight

ursus

I believe another switch, --alt-preset medium already uses -V3, though I'm not certain as to how else it differs from --alt-preset standard (maybe somebody else could jump in here). Try that and see if it meets your needs. If not, you might also try --alt-preset 192, though --alt-preset standard sounds better since it contains internal code tweaks not offered in the ABR settings.

Oh man......

If you think -Y loses too much, then just modify --lowpass to whatever suits you. Try 17.5.

Try "--alt-preset standard --lowpass 17.5 --ns-sfb21 4". It's a safer choice than adding -V3.

Thanks for everyone's feedback, this wasn't/isn't a troll thread

I really am interested in finding a near-equivalent to APS that is
maybe 10-15% tighter in bitrate ... not necessarily much lower,
either, since it seems once you get below say ~170kbps average,
we're no longer talking transparency or near-transparency ... if
future improvements related to 2-pass VBR, or nspsytune2, or
other improvements to LAME have this side effect all the better.
My ideal APS 'improved' would average around 192kbps rather
than the ~220+kbps which APS seems to favor. The idea is to
have r3mix-"like" bitrate efficiency while retaining say, 95% of the
transparency that APS provides. This way I don't need to retain
separate versions of my mp3s for listening on my main stereo
vs portable players.

I do know about --alt-preset-medium, but it's been pointed out
that that is simply a combination of command-line switches, and
lacks the code-level tweaks which Dibrom's presets employ. Also
it includes the -Y tweak, which I feel is too aggressive in throwing
out HF content (noticed this in listening tests, also corroborated
by comparing spectragraphs in CEP ... no I didn't just look at the
pretty graphs to come to this conclusion . I don't want to use
CBR or ABR as neither provides the quality that VBR can above
170kbps.

From reading some postings on lame-dev and here, I gather
changing the -V? setting alters the ATH curve for the entire mp3,
correct? As in, affects the encoding for all frequencies, not just
the above-16kHz bands as sfb21 tweaks do? It seems my ears
are still fairly good above 16kHz (maybe to 18kHz, or perhaps 19
but that's using tone sweep) and somehow the idea of whacking
HF detail above 16kHz aggressively seems just 'wrong' ... is this
thinking illogical? Why is it not better to gracefully degrade the
entire content -vs- encoding the lower frequencies at a high bit
rate and discarding [dis?]proportionate portion of HF content?
Maybe this is just one of those 'can't have it both ways' issues.

I will do some reading into the "sfb21 4" suggestion. Need to do
some testing to see whether I want to set the lowpass down at
17.5 ... Also what other changes does adding -V3 introduce other
than altering the ATH curve? The main thing that attracts me to
using APS-V3 is that it seems to smoothly bias the bitrate distro
around 192kbps (rather than 224kbps). APS-Y bitrate distro graph
just looks downright wierd, with obvious bias toward higher bit
rates, but the 224+ bitrate buckets are squashed VERY low ...

PS: why doesn't EncSpot report a known ABR when using plain
--alt-preset standard? If I add "-b128" to APS then EncSpot does
report ABR of 128 correctly. Is this just an EncSpot bug or does
APS forget to add this in the LAME tag?

ursus

Well, yeah, it would be better to gracefully roll off high frequency encoding, rather than simply pick 16KHz as a magical cutoff point. But because of scalefactor band 21 (sfb21), or rather the lack of scalefactor, 16KHz is a good cutoff from a bandwidth usage perspective. Basically, because of a limitation with the MP3 format, frequencies over 16KHz are much more "expensive" to encode at high levels of detail.

The sfb21 stuff, and how it relates to 16KHz+ frequencies, was explained by JohnV a while back, in case you haven't discovered this information yet.

The upcoming LAME 3.94 release should give us the ability to produce a well-tuned medium preset. It's difficult to do so with 3.90.2 because several low-level tweaking switches are not exposed in that release.

The -V switch modifies the position and the shape of the ATH curve and the lowpass value. This is why it is "risky" to modify APS with -V3. You are changing a lot of the thresholds and APS was tweaked by expecting a certain set of conditions.

Try preset --dm-medium.



I believe this is available with 3.93.1 or perhaps it is only available with alphas which I would NOT use for making mp3's besides testing.

Well, he's already said that he doesn't want to use the -Y switch, which is part of the packaged --preset medium (and the --dm presets have been gone for well over a year, heh). However, I should claify for ursus that --preset medium DOES include Dibrom's code-level tweaks (AFAIK) because it is based on --preset standard, but includes several command-line tweaks to decrease bitrate in the least offensive way.

Ursus, you have done blind ABX tests to verify that you can hear the effects of the -Y switch... right? (I have to ask, it's required by law here. )

I'll clarify this a little bit also...

If --preset medium is the same as it was when it was initially introduced, then it modifes the athadjust level, which directly affects some of the code level tweaks in --alt-preset standard. Basically what happens is that some of the tweaks will no longer kick in (rendering them a little useless). In essence, part of the way that the --preset medium reaches a lower bitrate than --preset standard is by using many of the quality tweaks less often. This happens to lower the bitrate, but it also almost assuredly lowers quality as well, since these thresholds were tuned very precisely with many listening tests. I don't think there has ever been any significant testing of --preset medium though to make sure that this rather haphazard solution actually doesn't impact quality significantly, which is the primary reason that I never cared too much for this preset, and why I don't personally recommend using it.

Perhaps they have been gone, but in 3.94a8 there is a preset called --preset dm-medium and a preset named --preset medium. Both have very different results both lower bitrate, but --preset dm-medium seems to be preset standard with a lowpass, while preset dm-medium has a higher bitrate preset medium does seem to include more high freqs. For example I will post a quick test I made earlier this week.


The test was more extensive than this including many more lines and a few other things besides bitrate but I posted this earlier in a testing thread and it still seems sufficent now.

Sorry, my bad. I misread your post, and I thought you were suggesting --preset dm-medium with 3.93.1 (the only version for which I tried to use the switch). I don't have any experience with the 3.94 alphas, but you reminded me that Dibrom had done some preliminary work on lower-fidelity Lame presets with the alpha.

And thanks for the info about --preset medium, Dibrom. Obviously I haven't tested it much, either!

You don't have to do any test to conclude that preset medium is lower quality than preset standard. That is obvious, and is the goal of this preset (reducing bitrate means reducing quality)

The point isn't whether it's lower in quality than --alt-preset standard. Everyone already understands that it should be simply due to it's nature.

The question is whether the tradeoffs made to lower quality are acceptable tradeoffs, and whether or not the quality degrades in a fairly gradual and linear fashion.

The real problem is that tradeoffs are valued differently by different listeners. I personally find limiting HF extension to be the best way to reduce MP3 bitrate while trying to maintain as much quality as possible. Sometimes increasing the sfb21 masking from the default 3.75dB to 7.75dB saves you 10%, 15%, even 20% in bitrate, without changing the lowpass. Of course, my ears become steeply insensitive to tones above 17KHz so I'm "biased".

After limiting high frequency retention, the next best step (IMO) is to play with the joint stereo switching thresholds. APS seems to use to many L/R frames and this can lead to substantial bloat. I'm saying this from a numbers perspective and not from a testing perspective. I'm sure Dibrom set the switching thresholds where they are for a reason, after on ABX'ing many samples. However, you can often find pretty substantial bitrate savings by loosening the switching threshold (thereby using less L/R frames) without changing the perceived imaging of a track.

Of course, there are many tracks that have neither extensive HF content nor wide stereo separation. For these situations, what I suggest above doesn't provide much bitrate savings. As a result, you have to start eating away the meat, such as by reducing the VBR quality level, and this is where problems typically arise. -V4 can be much noiser than -V2 yet going to -V4 is an almost guaranteed way to save about 25kbps in almost every file.

Linear and gradual is the way to go. You basically need to start with a "transparent" preset like standard and just loosen things a bit here and there until the resulting preset's appetite for bits is reduced to where you want it.

Thanks, I tried a variant of the above (lowpass 18 rather than 17.5)
and have been quite pleased with the results; while the bitrate graph
looks somewhat similar to 'APS -Y', listening tests of some material
esp. very some tracks with female vocal, with very complex instrumental
backing sound much cleaner vs. slightly dampened sound with 'APS -Y'.
I can't discern much difference between APS-Y and APS-lp18-nssfb21_4
on many tracks but for those that I can, the increased filesize is quite
reasonable for the improved clarity (and still at least 20kbps lower avg
bitrate vs. plain APS).

True to everyone's 'warning' about modifying APS preset with '-V3',
I did some more careful listening tests comparing 'APS-V3' against
'APS-lp18-nssfb21_4' and seems like the modified ATH curve DOES
have a negative effect on similar tracks as above (alto vocals, with
complex instrumental backing e.g. layered acoustic guitars -- example
some tracks from Tracy Chapman's TELLING STORIES album) ... the
effect noticed was the backing instrumentation would fade up and
down in perceived volume (I need to mentally focus on the backing
instruments and try to ignore the vocal portion -- as difficult as this
is with Chapman's arresting voice, as you might imagine -- since
-V? settings play with the ATH masking curve this should be expected
I guess (but wasn't expecting such noticible effects at V3!).

So bottom line, I'm going with "--alt-preset standard --lowpass 18
--nssfb21 4 -b128" in lieu of APS-Y or APS-V3 or preset medium for
my ~192 VBR needs ... I know I could probably lower the lowpass to
17.5 as my personal threshold seems just above 17kHz, but somehow
having that 500Hz buffer makes me feel better Thanks for everyone's
input on this topic!

Looking forward to more refined '--preset medium' in next LAME.

ursus