ABR vs VBR--where things stand

[pacohaas]

This was just added to LAME's USAGE file:

Using -V 5 or higher (lower quality) is NOT RECOMMENDED.
* *ABR will produce better results.

So it appears I will be reworking some of my guides for a single-CD divx.

[Huge]

Hmmm, interesting, but it doesn't tell you what ABR is better that -V 5?

I'm guessing something along the lines of 112?

[pacohaas]

well, I'd have to say that whatever average bitrate you get on any particular clip with -V5, you would have gotten a better mp3 using ABR at that average bitrate. I might start doing my mp3 encoding in 2 passes like that.

[Huge]

Did you come up with a decent 1CD setting using ABR?

[trg100]

Do you know why ABR should be better than VBR?

On paper the idea of letting the encoder choose whatever bitrate it thinks appropriate, without having to aim for a specific average, sounds optimal. Is ABR's superiority at lower bitrates a result of weaknesses in the VBR algorithm?

[pacohaas]

Originally posted by trg100
Do you know why ABR should be better than VBR?
no
On paper the idea of letting the encoder choose whatever bitrate it thinks appropriate, without having to aim for a specific average, sounds optimal.
yeah, that's what gets me too.
Is ABR's superiority at lower bitrates a result of weaknesses in the VBR algorithm?
it must be, though one thing I don't like is that ABR at low bitrates, <96 i believe, will automatically resample the output to 32000Hz. Though, I have done some listening tests between 48kHz, 44.1kHz, and 32kHz and the difference is not that major for files of equal size and with LAME's default lowpass. So, in short Huge, no I haven't come up with a decent 1CD setting using ABR, though 96 might not be a bad choice, try --dm-preset 96. Some movies are more compressable though(audio and/or video) which brings me to the conclusion that I am often drawn to: quality settings should be adjusted for each particular situation/movie/user.

[Gabriel_Bouvigne]

It's mainly because the current VBR scale is neither tuned neither heavily tested for high -V x values.

It doesn't mean that VBR couldn't achieve good results for low bitrates, but it means that our current settings are not very good for low bitrates.

About abr 96 and resampling: if you want you could use --abr 96 --resample 48 in order to keep the original sampling freq.

[pacohaas]

yeah, i know about --resample, but i'm wondering that since LAME automatically resamples when doing lower bitrates, should we then be resampling based on bitrates instead of half of us always saying "44100 is better" and the other half saying "Keep it at 48000"? It would seem that LAME takes into account that lower bitrates can't hold as much information so a resampling takes place. Perhaps we should have a new standard since we have the great tool SSRC to do our rate conversions, a single-CD rip should have 44.1 audio while a multi-CD rip should keep the 48(whether in mp3 or ac3)

[trg100]

About abr 96 and resampling: if you want you could use --abr 96 --resample 48 in order to keep the original sampling freq.

I read recently that doing this would just result in a lot of wasted bits i.e. the file would just be padded up to 48khz. Is this correct?

Also, it would be nice to think that lame downsamples for a good reason so we ought to let it get on with it.

BTW, I think the switch point for lame 32kHz resampling is <=102kbps.

Finally where can I get some (explanatory) info about SSRC options? Should I use dithering and what sort?

[pacohaas]

Originally posted by trg100
Finally where can I get some (explanatory) info about SSRC options? Should I use dithering and what sort?

There is none, other than the built in help. You'll have to do some testing if you want to know what they do. Actually there's some graphs and stuff on the webpage, but it doesn't really show much. Let us know if you find anything out. Personally, i just use none.

[MaTTeR]

Well after reading this thread I decided to encode a couple of movies with ABR and see how they compared to my VBR settings. I must say, I'm impressed. 7 out of 10 encodes were smaller when using ABR 112 and I couldn't hear a difference between quality. In the case of the crappy "Children of the Corn" DVD, the ABR MP3 actually sounded better than the crapola DD5.1 track. I'm sold

[pacohaas]

I've actually used this information in the opposite way, helping me find a good setting for backing up my CD collection to mp3. Since I'm no audiophile, --r3mix is a bit overkill for me, and ABR, as previously stated, has the same drawbacks as CBR, so I've decided to use -V4 in my backups.

[Ben]

Well, that's thrown all my expertise into the bin. I've always used low V settings (6-9) and for 1 cd rips I use V 8 with a max bitrate of 192, which sounds just great for me.

So I guess I should use ABR now??

[pacohaas]

Originally posted by Ben
Well, that's thrown all my expertise into the bin. I've always used low V settings (6-9) and for 1 cd rips I use V 8 with a max bitrate of 192, which sounds just great for me.

So I guess I should use ABR now??

it would seem so, what bitrates are you getting with these settings?

Also, there's really no point in using -B192 as there will only be a couple frames bigger than that at such high -V settings, the filesize savings are almost neglegable( < 0.3MB on a 2 hour mp3)

Originally posted by trg100 I read recently that doing this would just result in a lot of wasted bits i.e. the file would just be padded up to 48khz. Is this correct?

NO. MP3 as well as other compression methods will make no big difference whether you leave it as 48 Khz or downsample to 44.1 KHz. Size and quality will be very equal. Of course, for compatibility reasons i always downsample, SSRC will allow me to do this without any loss of quality ...

Also, it would be nice to think that lame downsamples for a good reason so we ought to let it get on with it. BTW, I think the switch point for lame 32kHz resampling is <=102kbps.

No. Yes. Lame's decision to downsample for bitrates under 103 kbps in ABR is not fully o.k., a 102 kbps ABR stream may still contain some information in the frequency range above 14 KHz without sounding really bad, and a 32 KHz sampling rate ( taking into account the necessary bandwidth for the antialiasing filters, even if being digital and high order ) is no good for frequencies above this. I personally would change this limit/behaviour of LAME and set it to 80 kbps or so ....

Finally where can I get some (explanatory) info about SSRC options? Should I use dithering and what sort?

Same here, no good explanation. As i understand it Shibata-San added this option to be able some inaudible very low level noise in order to avoid getting a number of consecutive blocks with digital zero ( absolute silence ) as this may disturb the resampling algorithm.

I try to come around this ( i hate the idea of adding something, even if being called inaudible ) by using 24 Bit mode for my SSRC conversion. You'll never ever have empty blocks ( digital zero ) in 24 bits mode i hope, but of course i dont know how Azid's AC3 decoding to 24 bits WAV is working exactly here ....

[dragonlz]

The reason why VBR does not YET give you optimal results is because the developers are still working on that part. I guess sometime in the future it would be great when they come up with a twopass mp3 encode algorithm.

Below is from lame usage file:
" VBR is currently under heavy development. Right now it can often result in too much compression. I do not recommend using VBR."

That explains why there are frames that are given really low bits during VBR encodings.

[Gabriel_Bouvigne]

This statement about vbr is totally outdated.
Vbr in highest modes is now really fine, and if you've got some low bitrate frames, it's mainly because those frames don't need more.
But yes, the fact that -V with high values is not so good comes from the fact that developpers didn't tuned those modes as much as the -V with low values.