Skip to main content ↑ ↓

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Vorbis absolute killer (Read 23300 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Vorbis absolute killer

Vorbis final version add noise in low bitrate. A lot of noise. On loud track, this particular behaviour is not clearly perceptible. But on quiet passage, this noise generation produce one of the worst flaw we can hear with audio coding.

VBR mode of Vorbis seems to be really accurate, except for low bitrate (I mean -q0). At -q 0, the bitrate can drop below 30 kbps with quiet passage (piano, quiet vocal, etc...), very common on classical music. In consequence, ABR ou CBR mode is safest : bitrate is closer to the target one, and quality is much, much  better (see, for exemple, the 64 kbps listening test on ff123 website, this summer, on the Liszt sample). Mischievous, I can say that vorbis is a pretty good codec on low bitrate with metal or pop music, but maybe the worst one with classical stuff. More seriously, I would say that vorbis is the less constant one in VBR mode at low bitrate with classical.

I encoded today a vocal works of Karlheinz Stockhausen : Stimmung. The beginning is in crescendo mode. Very, very quiet. So quiet that we need a very silent room to percieve the sound, or need to pump the volume up. A funny thing is to encode this with vorbis in VBR at the popular -q 0 mode. With this setting, the quality should be incredible, according to a lot of people. I must agree : incredibly... bad ! So catastrophic that lame, at 16 kbps, provide a better sound ! Noise, chirping, electric-like disturbance, etc... I let you judging it. And hope that Vorbis will correct it, because earlier vorbis codec were better than the final one. But is Vorbis able to be good with both loud and quiet music ?

Stimmung sample (30 seconds)

Vorbis absolute killer

Reply #1 –
Can you provide a link to a WAV version of the 30-second sample as well?

Vorbis absolute killer

Reply #2 –
Quote
Can you provide a link to a WAV version of the 30-second sample as well?

Just decode the ape sample I uploaded, and you will have the desired wav file
www.monkeysaudio.com

Vorbis absolute killer

Reply #3 –
And if you use Linux, you can most probably use the Monkey's Audio version Frank Klemm provides for Linux in order to decode it:
http://www.personal.uni-jena.de/~pfk/mpp/#ape

Although FLAC is the standard format for test samples here especially because of the cross-platform reasons..
Juha Laaksonheimo

Vorbis absolute killer

Reply #4 –
heh, I'm actually already equipped in the area of Monkey's Audio, but thanks anyway.  I'm a bit confused but I presume the above ape sample represents the uncompressed sound prior to a Vorbis encode?

Vorbis absolute killer

Reply #5 –
Quote
heh, I'm actually already equipped in the area of Monkey's Audio, but thanks anyway.  I'm a bit confused but I presume the above ape sample represents the uncompressed sound prior to a Vorbis encode?

Yes, you're right. So everybody can test with their own encoding.

Vorbis absolute killer

Reply #6 –
ah, thanks for clearing that up.

What this example represents is the only argument against VBR I can think of (at least for the lower bitrate range). It has been discussed here recently that soft passages suffer the most in terms of artifacts because the encoder is perceiving the quiet packages as possessing less audio information when, in reality, it's only the volume that is lacking.

You said the older Vorbis releases didn't possess this problem? I presume that it was there, just not as noticeable as it is now. 

Vorbis absolute killer

Reply #7 –
Quote
What this example represents is the only argument against VBR I can think of (at least for the lower bitrate range).


No, this doesn't represent an "argument against VBR", it simply emphasizes a flaw in the Vorbis encoder.  This is not a flaw in the theory, but rather a flaw in the implementation.

Quote
It has been discussed here recently that soft passages suffer the most in terms of artifacts because the encoder is perceiving the quiet packages as possessing less audio information when, in reality, it's only the volume that is lacking.


This is an overgeneralization.  The perceived volume of a sample actually has a lot of bearing as to how much masking there is, etc.  VBR coders with properly tuned psymodels can take advantage of this information to provide a higher quality/bitrate ratio.  Many high quality encoders are able to handle these situations fine and without creating problems.  Those that are not able to simply have a flawed or improperly tuned psymodel/bit allocation system.  Again, there's nothing wrong with VBR here, it's the fault of the encoder at hand.

In this case, the Vorbis encoder would simply appear to not be sensitive enough to low volume situations at the lower end of the quality scale.

Vorbis absolute killer

Reply #8 –
Quote
What this example represents is the only argument against VBR  (...)
You said the older Vorbis releases didn't possess this problem? I presume that it was there, just not as noticeable as it is now. 

1/ This is not an argument against VBR. Other VBR audio format are not concerned. Even Vorbis at higher bitrate (-q 2 approximately) doesn't suffer with such evidence with quiet passage. This is only a argument against universal superiority of VBR against CBR with vorbis -q 0 (and lower). VBR is very reliable with vorbis final version at high setting.


2/ Not exactly. This electric and noise & chirping problem occurs with older release. But with less intensity. But these old versions are really bad compared to the final one on loud music. Distorsion vs Noise. And noise is often prederable to distorsion.

Problem : can vorbis remove the noise issue without regressing in quality in other parts of the signal ?

Vorbis absolute killer

Reply #9 –
A strong lowpass (5000 hz) has positive effect : all chirps are removed. Pleasant at the beginning only

Vorbis absolute killer

Reply #10 –
1) I hear similar things with lame at 64kbs abr, it just starts a little later.

2) OK, so it is a bit rate misjudgement, but the file is half the size of what I got with lame.  It is a slipup
of the "set it and forget it" aspect of q value.  Can't you specify a minimum bit rate and still say q=0?

Vorbis absolute killer

Reply #11 –
Quote
I encoded today a vocal works of Karlheinz Stockhausen : Stimmung. The beginning is in crescendo mode. Very, very quiet. So quiet that we need a very silent room to percieve the sound, or need to pump the volume up.

But isn't this an unfair test setup? The codec makes use of the information, that the current portion of the piece has a very low (relative) volume and therefore differences are harder to detect for human hearing. But increasing the volume on playback destroys this relative difference (I assume that the work gets louder ), and exposes a defect which wouldn't be unusual disturbing (for this quality rate) when played on normal volume.

Try this: If you just want to listen to the very first minutes of the piece, increase the volume before encoding. This increases bitrate as well as perceived quality at the same ogg-quality setting.

But I agree: Vorbis has problems with low volume content (it seldomly reaches the nominal bitrate with classical music).

Vorbis absolute killer

Reply #12 –
Quote
But isn't this an unfair test setup? The codec makes use of the information, that the current portion of the piece has a very low (relative) volume and therefore differences are harder to detect for human hearing.

Yes and no.  Changes are usually judged as unnoticable because they are quiet cmpared to
something else in the track.  Relative volume should be relative to something.  On the other hand,
we are talking about q=0, so moral outrage over these artifacts may fall on deaf ears 

Vorbis absolute killer

Reply #13 –
Quote
But I agree: Vorbis has problems with low volume content (it seldomly reaches the nominal bitrate with classical music).

It has nothing to do with the volume, and everything to do with the deeper useful spectral depth of classical.  It hasn't been dynamically compressed from a 120 to a 50dB range to sound good on FM radio, and that's the only major difference (although that difference is pretty bloody major).  Any recording that hasn't seen a range compression will behave similarly.

Absolute volume has nothing to do with it.  Vorbis doesn't compute anything against an absolute volume (relative volume yes, absolute volume no, and relative volume is computed chronologically.  If the soft part is at the beginning.... Vorbis will assume that's max volume because that's the loudest it had seen until that point).

Monty

Vorbis absolute killer

Reply #14 –
Oh, I'd like to point out that APE samples are useless to me...  Not only do I have no windows machines, I don't run Linux on x86.  If there's no source, I can't use it.
FLAC and WAV for me, please...

(I'd like to test this sample... but I can't for now)

[update: got Klemm's Monkey stuff to compile.  I have the sample now]

Monty

Vorbis absolute killer

Reply #16 –
Thanks for comment.
I found another problem with -q 0 : noticeable distorsion on simple, quiet, piano music. Again, bitrate is going too low.
Of course, -q0 is only a 64 kbps nominal setting, and in consequence I can't expect a clean sound. Nevertheless, loud (compressed) music sounds in comparison really fine at the same setting. This contrast is annoying : one of the goal of VBR encoding is to provide a constant quality, regardless of musical genre, complexity or loudness. From this angle, vorbis fails with VBR at low bitrate : good alternative to mp3pro with loud music, catastrophic sound with quiet music. The actual codec isn't reliable on the 40-70 kbps range.

This new sample (flac 1.10 encoding) came from a recent digital recording. The composition is an extatic work of Franz Lizst, called Bénédiction de Dieu dans la solitude (Benediction of God in solitude). Length is 22 minutes. At -q0, the average bitrate of the whole track is only 27 kbps. Nice bitrate, but quality suffer too much. CBR 64 provide a less distord sound (final br is = 53 kbps), but introduce a terrible chirping ! Very strange...

Bénédiction (15 seconds).





And in this case, there is nothing unfair in this test : the loud passage are dominant (less than 3 minutes on 22'30 - 8% of the file) :

Vorbis absolute killer

Reply #17 –
very very interesting. I'm a relative noob in the realm of psychoacoustic codecs so don't take my comments too seriously.

As someone mentioned above, there is a similar behavior even with the LAME mp3 codec. How these codecs respond to quiet passages in low-bitrate (<100 kbps) VBR mode seems to give results that are less satisfactory than CBR.

Vorbis absolute killer

Reply #18 –
Quote
As someone mentioned above, there is a similar behavior even with the LAME mp3 codec. How these codecs respond to quiet passages in low-bitrate (<100 kbps) VBR mode seems to give results that are less satisfactory than CBR.

Lame can't be compared to other formats in VBR mode at low bitrate. For some reason I can't explain (psycho-accoustic ?), Lame VBR is pretty good at high bitrate, but can't avoid bad artifacts at low bitrate (under 120 kbps). So, in this area, it is better to use ABR. Safe too, because ABR is a good protection against some odd behaviour with low complexity music : the 22 minutes are encoded at 59 kbps with --alt-preset 64 (lame 3.93.1).


Remarks with the Liszt sample.

ogg -q0 = 24 kbps (bad distorsions)
ogg -q1 = 31 kbps (less distorsions, but significant)
ogg -q2 = 56 kbps (no distorsion, but an audible increase in background noise)

In consequence, lame --preset 64 seems to be less annoying than lame -q 2, and obviously, better than ogg VBR & CBR 64 kbps ! Can someone confirm my impressions, or am I victim of my personnal taste ?

Vorbis absolute killer

Reply #19 –
Quote
Absolute volume has nothing to do with it.  Vorbis doesn't compute anything against an absolute volume (relative volume yes, absolute volume no, and relative volume is computed chronologically.  If the soft part is at the beginning.... Vorbis will assume that's max volume because that's the loudest it had seen until that point).

I don't understand that. Increasing the volume by 15 db before encoding shouldn't affect the relative volume, but it causes a bitrate jump from 27 to 48 kbit/s with -q0.

Vorbis absolute killer

Reply #20 –
Quote
I don't understand that. Increasing the volume by 15 db before encoding shouldn't affect the relative volume, but it causes a bitrate jump from 27 to 48 kbit/s with -q0.

Oh, actually, there is one thing that is nailed to an absolute level; Vorbis won't let the ATH drift down to lower than -140dB absolute.

Vorbis absolute killer

Reply #21 –
Oh, and I added that sample to the 'tuning list that must be dealt with before next release'.  The very low bitrate support in 1.0 was meant to be a respectable but rough first cut, and it has obvious areas that require improvement.  I can point out a few you haven't, if you get really bored :-)

Vorbis absolute killer

Reply #22 –
Quote
Oh, and I added that sample to the 'tuning list that must be dealt with before next release'.  The very low bitrate support in 1.0 was meant to be a respectable but rough first cut, and it has obvious areas that require improvement.

Would be nice  Thanks, in advance.

Vorbis absolute killer

Reply #23 –
Quote
In consequence, lame --preset 64 seems to be less annoying than lame -q 2, and obviously, better than ogg VBR & CBR 64 kbps ! Can someone confirm my impressions, or am I victim of my personnal taste ?

I compared the following samples to the Liszt sample:
1. ogg -q0 = 24 kbps (really ugly)
2. ogg -q1 = 31 kbps (similarities to 1, but far better, smeared attack)
3. ogg -q2 = 56 kbps (better than 2, good noise)
4. ogg -q2.99 = 85 kbps (like 3)
5. ogg -q3 = 93 kbps (better, louder noise)
6. ogg cbr 64 = 52 kbps (disturbing "noise reduced" and "jumping" background noise, unsteady decay of sound (~tremolo))
7. lame -ap 64 = 61 kbps (damped, lower noise)
8. lame -ap cbr 64 (like 7)

So for this sample:
lame > ogg -q3 >> ogg -q2, 2.99 >> ogg cbr64 > ogg -q1 >> ogg -q0

The preference for Lame against ogg -q3 is a question of taste. Both have different problems with the sample, neither are extremely disturbing.

Vorbis absolute killer

Reply #24 –
Nice hierarchy. Thank you very much for your detailed reply