LAME 3.96.1's low bitrate cbr/abr not optimal, ABR/CBR regressing at lower bitrates? Options

[LoFiYo]

It seems to me that recently some people are starting to say/think that 3.96.1 is just as good as 3.90.x (I guess from their own tests and the inconclusive one here), but to me what is notably regressing is lower bitrate CBR/ABR tunings. I compared 3.96.1 and 3.90.3 in CBR96 mode (--[alt-]preset cbr 96) using mc_sich_short.wav. I have tested a bunch of other files also in even lower bitrates, and noticed similar results.

ABC/HR Version 0.9b, 30 August 2002
Testname: cbr96

1L = C:\sich\cbr96-3961.mp3.wav
2R = C:\sich\cbr96-3903.mp3.wav

---------------------------------------
General Comments:
ABXing #1 vs original was easy. #2 vs orig needed much more concentration.
---------------------------------------
1L File: C:\sich\cbr96-3961.mp3.wav
1L Rating: 3.0
1L Comment: Swishing/watery noise is obvious.
---------------------------------------
2R File: C:\sich\cbr96-3903.mp3.wav
2R Rating: 4.0
2R Comment: This is not bad at all. In noisy environments, it should sound really good. The distortion heard in #1 is mostly absent in this.
---------------------------------------
ABX Results:
Original vs C:\sich\cbr96-3961.mp3.wav
10 out of 10, pval < 0.001
Original vs C:\sich\cbr96-3903.mp3.wav
9 out of 10, pval = 0.011
C:\sich\cbr96-3961.mp3.wav vs C:\sich\cbr96-3903.mp3.wav
9 out of 11, pval = 0.033

If you use any old test sample, the difference is quite obvious. Of course, there probably are exceptions, but with my personal samples, 3.90.3 CBR/ABR presets (128kbps and lower) tend to be almost always superior to 3.96.1.

I understand that the recent development is concentrated on the lower VBR modes, and I am glad to see the development (Thank You!), but I just wanted to point this out here, because if 3.96.1 becomes the recommended version due to the growing popularity, people should know that there is a possible quality regression in some modes/settings like above, and there should be further testing around these bitrates in CBR/ABR presets.

Edit: for grammar & clarity

This post has been edited by LoFiYo: Oct 25 2004, 11:59

[[proxima]]

If you use any old test sample, the difference is quite obvious. Of course, there probably are exceptions, but with my personal samples, 3.90.3 CBR/ABR presets (128kbps and lower) tend to be almost always superior to 3.96.1.

I agree 100% with you. This is exactly what i found during my tests (some are available in the past threads) with --alt-preset 128. Watery artifatcs are noticeable.

[k.eight.a]

WOW! What can I say? There were some suggestions to use 3.96.1 for mid/low bitrates instead of 3.90.3...

[Dologan]

Bear in mind that these observations are for CBR/ABR. This doesn't mean that VBR for mid/low bitrates is showing flaws with respect to 3.90.3; from previous reports probably the opposite.

This post has been edited by Dologan: Oct 27 2004, 19:42

[LoFiYo]

Bear in mind that these observations are for CBR/ABR. This doesn't mean that VBR for mid/low bitrates is showing flaws with respect to 3.90.3; from previous reports probably the opposite.

I do think that for people who simply want to encode CD music, the mid to high bitrate VBR modes of 3.96.1 are more flexible than 3.90.3's VBR.

But surprisingly many people still either need to, or choose to, use low bit CBR for different reasons and for different purposes. So I still think it is important to maintain the quality of such basic performance of the encoder as CBR 128kbps (and lower). Depending on the type of music, 3.96.1's low bitrate CBR/ABR presets sound disturbingly untuned compared to 3.90.3.

[Dologan]

I do think that for people who simply want to encode CD music, the mid to high bitrate VBR modes of 3.96.1 are more flexible than 3.90.3's VBR.

But surprisingly many people still either need to, or choose to, use low bit CBR for different reasons and for different purposes. So I still think it is important to maintain the quality of such basic performance of the encoder as CBR 128kbps (and lower). Depending on the type of music, 3.96.1's low bitrate CBR/ABR presets sound disturbingly untuned compared to 3.90.3.

I am aware of this. This behaviour should be corrected before the newer LAMEs are to replace 3.90.3 at these settings. My comment was just meant to point out that the suggestions to use 3.96.1 at mid-bitrates referred to VBR modes. From k.eight.a's comment, it appears that your observation may be misinterpreted to mean that mid/low VBR in 3.96.1 wasn't as good as it appeared to be, when in fact it's got nothing to do.

[k.eight.a]

Well, I think that in general when we want to replace 3.90.3 recommended compile the newer one must be at least as good as the previous. And according to the first post 3.96.1 is not that one. This is my observation. And my previous post was ment to be... Sorry for the confusion...

This post has been edited by k.eight.a: Oct 27 2004, 21:55

[dev0]

WOW! What can I say? There were some suggestions to use 3.96.1 for mid/low bitrates instead of 3.90.3...

The recommendation was for 3.96's low/mid bitrates VBR modes.

I'm aware of the quality regression in ABR/CBR. [proxima] first brought it up when the initial 3.96 vs. 3.90.3 test started AFAIR.
I think Gabriel is aware of the problem too as he started working on some tunings and created a thread calling for testing (which got little recognition by the 3.96 loving public).

[shadowking]

A few tests so far on 3.96.1 and I've found several ugly warbling,ringing artifacts with abr 128, 144 - but on 160 they become much weaker. Some samples were solved entirely by using V5. Some samples still had these problems on v5 and to a lesser extent v4 and abr 160. APS solved them. I didn't test 3.90.3.

V5 is great for 130k and that's what i use for the portable - most music seems transparent but the hard samples might be audiable if you are familiar with them. V4 or abr160 would be transparent for casual listening and I don't think i'd hear anything that would bother me outside of abxing.

[Gabriel]

I'd like some pointer to samples and specific locations where you found 3.96.1 cbr/abr to be clearly inferior to 3.90.3.
I tryed to abx mc_sich_short with 3.90.3 and 3.96.1 using 96kbps cbr, but failed. On my laptop, the integrated chipset is really bad. It would help me if you could point to some tracks, but also mention specifically where in the track the difference is obvious.

note: a correct external soundcard would probably help me a lot. Someone got a whole new computer, perhaps I could have an external soundcard?

[LoFiYo]

I'd like some pointer to samples and specific locations where you found 3.96.1 cbr/abr to be clearly inferior to 3.90.3.
I tryed to abx mc_sich_short with 3.90.3 and 3.96.1 using 96kbps cbr, but failed. On my laptop, the integrated chipset is really bad. It would help me if you could point to some tracks, but also mention specifically where in the track the difference is obvious.

note: a correct external soundcard would probably help me a lot. Someone got a whole new computer, perhaps I could have an external soundcard?

Wow, that's gotta be a tough environment for tuning an MP3 encoder!

As I said above, providing that you have a fairly decent system and a set of headphones, if you use any of the test samples available around here, I think you can hear the difference at most times. I thought mc_sich_short was obvious enough, but I will try to find something even more obvious (to me) later this week or next week.

My uneducated guess is that the problem with 3.96.1's low bitrate CBR/ABR presets is that they use too high lowpass for low bitrates as compared to 3.90.3, and it's showing.

Thank you very much for trying to work things out

[Gabriel]

I thought mc_sich_short was obvious enough, but I will try to find something even more obvious (to me) later this week or next week.

If the distortion is more obvious at a specific location in this clip, it would help if you could mention this location.

Regarding the lowpass, I do not think it is the culprit. At 96kbps, there is only a 100Hz difference between 3.90.3 and 3.96.

[LoFiYo]

Gabriel,
Please try hihat.wav in 96kbps CBR. To me, the pre-echo artifact of 3.90.3 (the first 3 hits especially) is more acceptable than that of 3.96.1.

[LoFiYo]

If the distortion is more obvious at a specific location in this clip, it would help if you could mention this location.

The distortion in mc_sich_short.wav I noticed was the first 2 seconds and around the 6th sec to the end. The distortion sounds the same to me with 3.96.1 and 3.97a4.

And I think I found a quick & dirty solution. Add --resample 32. One of the differences between 3.90.x's "--alt-preset cbr 96" and 3.96.1 (& 3.97a4)'s "--preset cbr 96" is the output sampling frequency. With 32kHz output sampling frequency, the artifact mostly vanishes with this sample, and the quality become roughly the same with all 3 versions.

With hihat.wav, 3.90.3 is still noticeably better than 3.96.1 and 3.97a4 even with --resample 32 although it does reduce the pre-echo artifact a little bit (Please see below).

ABC/HR Version 0.9b, 30 August 2002
Testname: hihat 3.97a4 --preset cbr 96 --resample 32 VS 3.90.3 --alt-preset cbr 96

1L = C:\hihat\3.97a4-cbr96-32khz.mp3.wav
2R = C:\hihat\3.90.3-cbr96.mp3.wav

---------------------------------------
General Comments:

---------------------------------------
1L File: C:\hihat\3.97a4-cbr96-32khz.mp3.wav
1L Rating: 1.0
1L Comment: The third and fourth hits are especially bad.
---------------------------------------
2R File: C:\hihat\3.90.3-cbr96.mp3.wav
2R Rating: 2.3
2R Comment: Distorted, but not as annoying as #1.
---------------------------------------
ABX Results:
Original vs C:\hihat\3.97a4-cbr96-32khz.mp3.wav
10 out of 10, pval < 0.001
Original vs C:\hihat\3.90.3-cbr96.mp3.wav
10 out of 10, pval < 0.001
C:\hihat\3.97a4-cbr96-32khz.mp3.wav vs C:\hihat\3.90.3-cbr96.mp3.wav
10 out of 10, pval < 0.001

This post has been edited by LoFiYo: Nov 21 2004, 05:40

[Gabriel]

Would you mind trying "--preset cbr 96 --lowpass 15100" ?

[LoFiYo]

Would you mind trying "--preset cbr 96 --lowpass 15100" ?

That setting sounded pretty much the same as when I added --resample 32. 3.90.3 seems to create smear instead of pre-echo, and that makes it sound less bad to me than 3.97a4. Please see below.
________________________________
ABC/HR Version 0.9b, 30 August 2002
Testname: hihat.wav (3.90.3 --alt-preset cbr 96 VS 3.97a4 --preset cbr 96 --lowpass 15100)

1R = C:\hihat\3903-cbr96.mp3.wav
2R = C:\hihat\397a4-cbr96-lp15100.mp3.wav

---------------------------------------
General Comments:

---------------------------------------
1R File: C:\hihat\3903-cbr96.mp3.wav
1R Rating: 2.2
1R Comment: The distortion sounds to me closer to smear than pre-echo. It is less disturbing than #2.
---------------------------------------
2R File: C:\hihat\397a4-cbr96-lp15100.mp3.wav
2R Rating: 1.1
2R Comment: The 3rd and 4th hits are noticeably affected by pre-echo.
---------------------------------------
ABX Results:
Original vs C:\hihat\3903-cbr96.mp3.wav
10 out of 10, pval < 0.001
Original vs C:\hihat\397a4-cbr96-lp15100.mp3.wav
10 out of 10, pval < 0.001
C:\hihat\3903-cbr96.mp3.wav vs C:\hihat\397a4-cbr96-lp15100.mp3.wav
10 out of 10, pval < 0.001

OT, but I am impressed by the performance of 3.96.1/3.97a4's "-V 7" for casual listening!!

[ezra2323]

I still find this situation confusing. If you are making files for a flash player (or some other use of 128 sized MP3) what would you use? I use APS but still find the occasional use for a 128 sized file.

Please rank the following versions/presets in order, 1 being the best, with regard to sound quality only.

3.96.1 -V5
3.96.1 128 (ABR)
3.96.1 CBR 128
3.90.3 128 (ABR)
3.90.3 CBR 128
Other version/preset that produces 128 sized files

[mithrandir]

I would select 3.96.1 -V5 and 3.90.3 128 (ABR) as the best choices.

[ezra2323]

Is it just "-V5" or is there some additional command as well? (such as was used in the multi-format test)

[Gabriel]

If you are sensitive to high freq problems and are hearing problems with V5 on your flash player, you could try to use the same additionnal switches as in the multiformat test.

But for real use on a small flash player, you might also try V6.

[ezra2323]

If you are sensitive to high freq problems and are hearing problems with V5 on your flash player, you could try to use the same additionnal switches as in the multiformat test.

Actually, I am. I notice most artifacts on cymbals and hi-hats. I rarely can ABX vocals, bass, lower guitar chords, and piano from 128 to 320.

Will the additional switch from the test further inflate the file size? I notice V5 already averages around 135-140 from the music I have been encoding.

[Gabriel]

Will the additional switch from the test further inflate the file size?

Yes, by up to 5kbps

[ezra2323]

Thanks, Gabriel.

Hopefully this will help. Some songs sound great but others just make me cringe when the cymbals smear.