[hifou]

I am wondering: If you take a WAV source and convert it to Vorbis in Q9 (say at 320constant), then convert back to Wav from the Vorbis and convert back to Q9 from this Vorbis what happens ?
 
I know that dual conversion looses more data, but is this really true ? My reasoning is that the first WAV->Vorbis conversion will remove  whatever part of the (inaudible) spectrum it decides to remove and perhaps smooth a little the transients.
 
But then, if you convert once more this already simplified "spectrum" I could think that there will be much less stuff to throw away comparing to the first conversion.
 
I'm not sure if I express myself well, but say you have a loud symphony orchestra and a "weak" flute in front which gets "optimized out". It's share of bits get allocated to the rest of the audible and louder spectrum and we call it a day.
 
If you re-optimize this file again, the flute is already optimized so there's nothing to remove more.
 
Have you tried actually to do this and then ABX the results when using *high* bitrate codecs ?

 

[VNandor]

Theoretically, there could be.

   
But then, if you convert once more this already simplified "spectrum" I could think that there will be much less stuff to throw away comparing to the first conversion.
 

Since lossy encoding will *always* remove data it doesn't care how much there is left to throw away, it always will.
 
I've done a lot blind testing regarding (high quality) lossy vs. lossless and I've failed nearly all of them, except when I tested unusually hard to encode samples. Based on that experience ,I would say you would have to reencode the sample for quite a few times before the difference becomes noticeable compared to the original wav file and comparing it to the first lossy version would be even harder. So in practice it most likely won't be noticeable if you did it two times only.

 

[hifou]

I agree there will be loss, but we're not talking about preserving the exact original waveform but more of decomposing it in its harmonics and then analyzing it in the transform domain where most of the inaudible parts will have been shaved or already degraded. That is they will already be using whatever low bit count quantization was chosen at the beginning.
 
In other words, I think they will have been "aligned" to a lower precision and a second conversion is likely to leave them where they are, with the same lower precision and at the same "level" as the jump to a new level would be too big.
 
This is why my idea is that the first conversion is the major one and a second will introduce less artifacts. Then on specific material this might not be the case.
 
I think I need to do my homework and my ABX tests for a definitive answer.

 

[Don Hills]

It's easy enough to test... encode some music. Subtract it from the original and check the level of the residual that was left out by the encoder. (But remember of course that just because it was left out, doesn't mean you can hear its absence.) Then encode a second time and subtract the second encode from the first encode. The residual will indicate how much was left out by the second encode.

 

[spruce music]

  It's easy enough to test... encode some music. Subtract it from the original and check the level of the residual that was left out by the encoder. (But remember of course that just because it was left out, doesn't mean you can hear its absence.) Then encode a second time and subtract the second encode from the first encode. The residual will indicate how much was left out by the second encode.

Good suggestion.  Just tried it with some John Mayall music.  Used a medium quality setting on the Ogg.  You could every now and then hear a very faint guitar chord and hint of voice.  Otherwise mostly cymbal swishing and general watery noise sounds. The difference in the first and second encode was just about the same as between original and first encode. So between original and second encode the same kinds of residuals just at twice the level.  Voices just now and again were more than a hint and clearly heard.
 
So compared to the original you would hear some difference if you listened very carefully.  Yet at the same time, the 2nd encode on its own had no horribly nasty artefacts.  I doubt anyone simply listening would comment about anything being wrong with it.

 

[mer2329]

Yes.
Simply put you are converting a lossless format to a lossy format.
Now do note the loss may not be noticeable. But there is loss.

Sent from my LG-H811 using Tapatalk

 

[hifou]

I did a little more testing. I did a FLAC (1) -> VORBIS (2) -> VORBIS (3) conversion with Q9 on vorbis format. The resulting waveform from subtracting vorbis-2 and vorbis-3 is noticeable. If you try to listen to it you actually hear what was left out. Mainly transients and high part of the spectrum. This makes sense.
 
I thought I could easily spot the differences between the files. Armed with a good pair of iems and full size hp I did a proper ABX test in Foobar2000. Well, I failed, miserably. Whatever I thought I was hearing it must have been in my mind.
 
My score was Ok at the beginning but then it fell quickly. Might have been listening fatigue or not the proper music. 
 
For the purpose of enjoying the music, it's completely transparent.
 
I guess this answers my question.