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 flaw with classical instruments? (Read 16174 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Vorbis flaw with classical instruments?

I'm currently comparing different versions of Vorbis as a pre-trial before the new 64kbps test. I must say 64kbps sounds impressive these days.
Still Vorbis has one giant flaw, that makes some samples horrible and I can't see that it would take many extra bits to get it right.

The effect I speak of is a vibrato/vibration effect. On the clip "Carbonelli.wv" provided by Guruboolez it's very easy to spot.
http://www.hydrogenaudio.org/forums/index....pe=post&id=1437

The violin sounds like it has a cold, rusty. Almost like flutter effect in tape recorders. I'm only guessing but it's like the pitch keeps bobbing up and down.
perhaps it hits between two fft bin's ? This is very irritating to listen to, and ruins the track completely. All other typical vorbis artifacts are not annoying to me at this bitrate.

Vorbis flaw with classical instruments?

Reply #1 –
Which Vorbis version did you examine?

Vorbis flaw with classical instruments?

Reply #2 –
xiph 1.0.1
xiph 1.1.0
archer_RC4
aoTuv PB4

They all sound the same to me with regard to this artifact

Vorbis flaw with classical instruments?

Reply #3 –
This sample is difficult in this bit rate. Although, as for Vorbis, volume shakes, probably, a heavy incidental sound is added in other codecs.

If it says about aoTuV, I will think that it becomes a little better by formal beta4.

Vorbis flaw with classical instruments?

Reply #4 –
Quote
The effect I speak of is a vibrato/vibration effect.
[a href="index.php?act=findpost&pid=291258"][{POST_SNAPBACK}][/a]


I agree with you. But this problem is not specific to vorbis; you can find it (or at least similar distortions) with many lossy formats at low- and even mid-bitrate.

To complete the answer, I must say that the performance of vorbis at 64 kbps are often very poor for my taste when I test with classical music samples (first thing I usually do). It's only when I switch to another category of sample that I find vorbis really decent at 64 kbps. The latest development of aoTuV doesn't seriously change the low performance of vorbis at low bitrate with classical music [not all samples/music of course]. It's probably the main reason that coud explain my current lack of motivation for testing it. On the few comparison between SVN & aoTuV I did, both were equally poor, without noticeable improvment.

N.B. The Carbonelli.wav sample reveals a typical issue of lossy encoders with this kind of sample (highly tonal, without attacks). Other encoders (I mean SBR one like HE-AAC) have also severe problems (ringing). That's why I proposed it to Sebastian Mares for his listening test: there's not only pre-echo

Vorbis flaw with classical instruments?

Reply #5 –
As I have some knowledge about audio/video compression I wonder if somebody could explain why the violin is so hard to compress.

It would seem to me the spectrum of the instrument is either constant or periodic in nature, making it easy to predict.
Is it a subsample/alias problem due to lack of time resolution ? then shorter blocks should help.
What if one combines the time information from shorter fft blocks or wavelets. to modulate on top of the regular block length vorbis chooses for this case, to prevent the aliasing.
One could use a simple periodic predictor so the bit overhead is minimal for this type of instrument. This would  have to be implemented in the decoder too ofcourse....

Perhaps the ear is more sensitive when the note is so constant. A encode/decode loop for "steady" tones could be made to model the amplitude modulation that are close to the block length. Increasing bitrate, shorten or lengthen block size to reduce the effect. In other words make an addition to the psy model to also take this effect into consideration, and force it to be below a certain treshold. This would not change the decoder

I wonder if I make any sense at all 

Vorbis flaw with classical instruments?

Reply #6 –
Quote
Perhaps the ear is more sensitive when the note is so constant


Do you know the frequency range of violin?

some information about ears:
1. very sensitive to 3 - 4 kHz
2. become less and less sensitive when the frequency goes down from 3 Khz to 20 Hz,
    this is also true for 3 kHz to 10 kHz
3. very insensitive to frequencies above 10 kHz
4. can still detect frequency below 20Hz or above 20 kHz if it is louder than 120 dB


maybe violin is in 3-4 kHz range?

Vorbis flaw with classical instruments?

Reply #7 –
I think, from rough estimations, the range of a violin should be from around 500 ~ 6,500 Hz, although the top 1kHz I don't think would be used all that much. Has anybody any better idea? I'm working on the assumption that the bottom note is just above tuning 440Hz and there's about a three and a half octave range (although it's probably a bit higher - the 'cello at least can get just over 4 octaves).

Vorbis flaw with classical instruments?

Reply #8 –
I'd describe the effect/artefact as a very quick tremolo (independant for each sinosodial) which could be explained due to too heavy undithered quantization of the MDCT samples. Sometimes energy/loudness of a harmonic is dropped (MDCT coeff has been rounded towards zero) and on some packets energy/loudness of a harmonic is increased (MDCT coeff has been rounded towards +/- infinity). This quick amplitude modulation creates artificial partials and makes the sound dirty.

I can't think of a work-around other than increasing the SNR for tonals (which of course increases bitrate)


my two cents,
SebastianG

Vorbis flaw with classical instruments?

Reply #9 –
Quote
It would seem to me the spectrum of the instrument is either constant or periodic in nature, making it easy to predict.
Is it a subsample/alias problem due to lack of time resolution ? then shorter blocks should help.
[a href="index.php?act=findpost&pid=291290"][{POST_SNAPBACK}][/a]


Vorbis doesn't make use of any kind of signal prediction (like TNS or Prediction of AAC main profile). It's a transform-only codec (considering MDCT samples only - the floor1 curve is coded via spectral prediction). Therefore I guess that longer blocks coupled with a higher SNR would lead to a slight improvement in this case.  (longer block = better energy compaction for steady signals --> kind of a prediction substitute)


SebastianG

edit: fixed typo & added info about floor1 prediction

PS: Plain scalar quantization sucks anyways for low SNRs.

Vorbis flaw with classical instruments?

Reply #10 –
When the artifact is this obvious the psy model must be flawed somehow. If the variations in the violin had been quicker the sidebands created around
each harmonic would be taken into consideration. However when the modulation sidebands come very close to the harmonics, meaning the period approaches
the block length in duration it seems the psy model no longer sees them, and don't allocate bits properly.

The fact that the artifact sound so periodic, suggests that it is not rounding errors or other more or less "random" fenomena. A periodic effect should be easy to encode or correct for without using up  alot of bits. Noise like artifacts however is troublesome, and costs alot of bits to correct. I say all  this on general terms, and partially guessing as my background does not come from working with vorbis or similar codecs.

if somebody has time ( aoyumi?  )  they could try different block lengths to see if it affects the problem. perhaps very large block lengths would actually work
better for this case, but maybe the masking model needs to be modified for frequencies  close the the dominant ones. perhaps under 20hz or something where the ear no longer hear things as tones, but modulation on top of the tones.... I'm just guessing here

Vorbis flaw with classical instruments?

Reply #11 –
Quote
When the artifact is this obvious the psy model must be flawed somehow.
[a href="index.php?act=findpost&pid=291349"][{POST_SNAPBACK}][/a]

heheh  What do you expect ? perfect transparent sound at 64 kbps ?
Let's consider that the Vorbis encoder tries to encode a bandwidth of 16 kHz at q0 (I think it does). 64 kbps for a stereo signal would then translate to 1 bit per sample per channel on average. Not very much to encode harmonics without noticable amplitude modulation artefacts.

It's like "lame -b128 -k"



SebastianG

Vorbis flaw with classical instruments?

Reply #12 –
I know it's only 64kbps

I for one would be happy with more stereo colapse and muddy dirty sound if this tonal vibrato effect could be almost removed.
It's the only artifact vorbis has that makes tracks really annoying at this bitrate, atleast to my ears.
Perhaps others thinks this is the right balance between artifacts. please speak up

I just want to help out to make Vorbis as good as possible

Vorbis flaw with classical instruments?

Reply #13 –
for fun I tested higher bitrates with a constant 16KHz lowpass.

I start hearing it at q2 = 96kbps but I could live with this amount.  at q3 = 112kbps it's gone

Vorbis flaw with classical instruments?

Reply #14 –
- comment about point-stereo removed -

Vorbis flaw with classical instruments?

Reply #15 –
SebastianG you might be more right than you know.

I recoded this track in mono and q=-1 (approx 32kbps) , and the artifacts is gone!?!
it's gone even at q=-2 giving 24kbps !
So for fun I used  cooledit and encoded left and right channel by themselves at q-1 with aoTuv PB4 then joined them back together after decoding them to wav again.
And what do we get ?!  a good sounding stereo Vorbis encode of carbonelli at 64kbps
And this without getting the benefit of channel similarities, showing that Vorbis has a great possibility for further improvement here.

I put the files together in a zip:        http://www.omegav.ntnu.no/~kamben/carbonel..._stereo_tst.zip

Perhaps Guruboolez could try this on his collection of "killer" tracks

Vorbis flaw with classical instruments?

Reply #16 –
Interesting
Hope this will help developers to correct some obvious issues with current encoder at low & sometimes mid-bitrate

Vorbis flaw with classical instruments?

Reply #17 –
Quote
Let's consider that the Vorbis encoder tries to encode a bandwidth of 16 kHz at q0 (I think it does). 64 kbps for a stereo signal would then translate to 1 bit per sample per channel on average. Not very much to encode harmonics without noticable amplitude modulation artefacts.


Now is pre-echo related to the Gibbs Phenomenon in general? I just stumbled upon it and what exactly is the correlation between the two?.
budding I.T professional

Vorbis flaw with classical instruments?

Reply #18 –
Quote
SebastianG you might be more right than you know.

I recoded this track in mono and q=-1 (approx 32kbps) , and the artifacts is gone!?!
it's gone even at q=-2 giving 24kbps !
[a href="index.php?act=findpost&pid=291666"][{POST_SNAPBACK}][/a]

That's exactly what I tried after posting the point-stereo comment. But I cannot confirm your results (OggEnc 2.3, libVosbir 1.1.0). The tremolo of the harmonics is still there (check -q0). Though, it's slower and a bit less strong for -q-1 (and below) due to the larger blocksize).

I think the improvement you noticed is due to the larger blocksize and isn't related to point-stereo.


SebastianG

Edit: well! I guess it IS also related to point-stereo. Now, i feel like the artefact is less strong for the mono files. 

Vorbis flaw with classical instruments?

Reply #19 –
Could someone post something lighter (5 MB for evaluating a 15' sec sample encoded at 64 kbps is a bit much for my poor dial connection)? Thanks

Vorbis flaw with classical instruments?

Reply #20 –
http://www.omegav.ntnu.no/~kamben/carbonel...o_tst_LIGHT.zip

wav files have been encoded with musepack --extreme.

you can decode the left and right ogg files yourself and put them into a stereo wav using cooledit or similar program.
Musepack recode might be slightly more metallic but I might just be "hearing things"
Atleast it adds no vibrato effect

Vorbis flaw with classical instruments?

Reply #21 –
Thank you  A single lossless encoding of the reconstructed file would be perfect, but the mpc file is probably enough.

I've listened to it, and despite remaining tremolo, I agree that the sound is here much better. Not perfect of course, but the artifact is clearly more acceptable than current SVN/aoTuV encodings at -q0.

Vorbis flaw with classical instruments?

Reply #22 –
Another track where this effect is very obvious:
http://www.omegav.ntnu.no/~kamben/Kansas__...n_The_Wind.flac

Also tried dual mono on another track that suffered from something I'd call a stereo wobble effect. And ofcourse it went away with dual mono.
The interesting thing was that to my ears the dualmono 2x32kbps sounded very close to the stereo 64kbps encode but without the slight bakground wobble.
This supports my growing feeling that both the stereo wobble and the vibrato effect are two sides of the same thing and is related to the handling of stereo.
And also that it would be possible to reduce these effects without eating up alot of bits.
I read a little about Vorbis stereo modes and it seems very complex, but also very powerfull.
However it seems the current versions are not optimal , and that these effects could be greatly reduced with some tuning and give Vorbis a nice quality boost at this bitrate.

Vorbis flaw with classical instruments?

Reply #23 –
Quote
I read a little about Vorbis stereo modes and it seems very complex, but also very powerfull.


It is. It's based upon square polar representation.  Very unusual to use with in conjunction with VQ.  I think this were Vorbis get's most of it's power from.  The bit savings is substantial from what I read in the past.  I don't know how it would be possible to tune this without actually removing it all together.  Two years from now somebody is going to be complaining about the same thing they were two years ago  I have heard the artifact you are referring to. I listen to a streaming classical Vorbis station at -q -1. It's not that noticeable only once and a while do you hear the problem with decaying harmonics that Sebastian was referring to. I had just always been convinced that these were due to block sizes even though Vorbis uses large block size for -q -1.
budding I.T professional

Vorbis flaw with classical instruments?

Reply #24 –
well if nobody wants to improve it, then it will stay the same...

what about using phase stereo ?

if this effect isn't covered by the normal psy models, or it's hard to  tune based on weights and normal criteria, then one must make an alogrithm
that can measure this effect and choose stereo mode/quantisation based on it. My impression from reading post by Guruboolez and my own testing is that this happens abit too often. 
On some tracks it dominates completely, and then it's of little use that the transients and fidelity otherwise are great.