Can Ogg, in theory, be tuned to beat MPC --standard?
Also, will bitpeeling (once stable) allow for higher bitrate ogg files encoded with current binaries to be later reduced in size as future binaries improve, with little to no perceptual loss of quality?

Probably not. Ogg Vorbis, as a pure transform coder, will always have poorer time resolution and thus cannot handle sharp transients as well as MPC which is a subband coder.

I'm not sure this is necessarily true; it seems that good enough transient detection would render the point moot, as you'd switch to short enough blocks to get whatever time resolution was necessary to be transparent on that segment of music. The problem is likely that transient detection isn't currently good enough (or, alternately, that short blocks aren't quite short enough).

That's true. Reading through the Vorbis I specs, I do remember seeing how the frame sizes can vary between 64 and 8192 samples as powers of two.

But transform coding is a rather limited concept now since its been with us for so long and everything that can be done to improve has probably been mostly done. Subband and wavelet coders, with their joint time-frequency representations, provide more avenues of improvement. In a lecture by Martin Vetterli, I heard that with the case of wavelet coders, non-linear approximation (choosing coefficients based on magnitude rather than order) results in better distortion reduction than transform coding. That sort of convinced me that transform coding and working in the frequency domain, in general, has lost its sparkle.

But of course, reliable transient detection certain covers the pitfalls of Ogg Vorbis and it is worth looking into.

Also I should point out that Vorbis II will include wavelets? So Ogg Vorbis may be able to match MPC, after all.

Yeah, pretty much like subband coding. Oh, wait...

Well, being from the image coding area, it seems to me that transform coding is practically dead. No-one is interested in transform coding or improvements to it anymore since wavelet coders and their complicated renditions outperform them.

It mostly explains why ICASSP 2003 didnt accept my paper which was (surprise surprise) on improving transform coders.

But then, wavelet coders havent had an impact on audio yet. I'm waiting to see an EZW or SPIHT coder for audio.

@Garf: So do you think vorbis can be tuned to be better than musepack --quality 5?

I think transform coding still has more potential and room for improvement than subband coding, but it's not easy, and I don't see much developement going in that direction.

@QuantumKnot

Where did you find the Vorbis II specs? Do you have any additional information about it?

Practically speaking,... which types of music will the aspects you'll discussed,
about the encoders,...effect ?
Or, how will it affect the music.

for eg


Would like to know especially in relation to Ogg Vorbis.

Thanks.
~~~~~~~~~
Autumnrain

I sometimes prefer the poorer time resolution in low-bitrate coding over subbanding the time domain data instead for better frequency resolution. There is going to be subjective quality lost no matter which way you split it. With Vorbis the loss is at least graceful as compared to something like MP3 where you have to hear the quantization noise spread out over the frame which get's really annoying after a while or WMA with the "metallic artifacts". At the least the block boundary artifacts in transform coders are tolerable. Besides the masking curves can always be readjusted to deal with sharp transients for the time being. Multiresolution anaylsis might be the next best step for Vorbis whenever that will be (Wavelets). Better time localization. If you have golden ears though your best bet is to probably stick with MPC especially in 200 kbps+ range for those pathogenic samples.

tangent:

For audio coding, I suspect transform coding is here to stay. Improvements are always at the horizon but the same applies to subband coding I guess.

fragtal:

I read it in the Vorbis specs. It says that Vorbis I uses MDCT and Vorbis II will use wavelets.

AutumnRain:

I'm not much of an audio coding person, since I come from an image coding background. So I'm only talking about it from a general signal processing perspective. Fourier-based transform coding involves representing your signal using sum of continuous waves like sines and cosines. They stretch forever in time but have a certain definite frequency (sort of like the tone or pitch). Hence transform coders are always good at accurately representing frequencies in your music as there is quite compact support in the frequency domain. But sine and cosine waves stretch forever in time and so they suck a bit when representing transients in music like sharp attacks. I mean, if you got some silence and then intermittently some sharp attacks which occur at specific time periods, then transform coders require (theoretically) a tremendously large amount of information in order to reconstruct that sharp attack at that specific time accurately. With coarse quantisation errors that will occur in high frequencies, these translate to the distortion that occurs throughout the entire frame and most of you would know this as 'pre-echo'. Thus in order to reduce the 'inability' of a transform coder to compress this, we reduce the frame size in order to make the signal more local in time which corresponds to increase in bitrate, so to speak.

With wavelet coders (I'll refrain from speaking from the subband coding perspective), the idea is to represent your signal (audio) as a sum of weighted basis signals which are compact in time (not extending forever like sine wave) as well as frequency. They are sort of an 'in-between' method I guess. The low frequency part of the audio (the trends) are decomposed into weighted scaling functions while the high frequency residue parts (the transients) are represented by weighted wavelet functions. Now when one looks at a wavelet, they will notice just how compact it is in time and thus they were very well suited to representing sharp transients accurately. As Martin Vetterli said in a recent lecture of his, wavelets function as 'singularity detectors'. Hence there will be less information that is needed to represent that sharp peak and less information (wavelet coefficients) means less bitrate to get the same fidelity. This is the single biggest advantage of subband/wavelet coders over transform coders, which is why I came to the view that Vorbis will never be as good as MPC. Even though you can add in intelligent transient detection in Vorbis to reduce the frame size and code transients better, what you gain in fidelity you will lose in bitrates which is why Vorbis files blow out on tracks which have lots of sharp transients. So even if Vorbis matches MPC, which is very likely, it will surely lose out in terms of compact file size. It is an intrinsic property of transform coding sadly.

Hopefully I explained that as simply as possible and as accurately as I can. Please feel free to correct mistakes in explanation or concepts if you notice them.

Hi.

QuantumKnot & HotshotGG....that was very informative.... thanks.

@QuantumKnot -
I'm not much of an audio coding person, since........
Hey :-) I don't know much technically either...less than you in any case.
But that stuff was interesting....since I 've always sought off /?...favoured OGG

Also, what would be the effect of what you'll are speaking about , generally/broadly speaking;
in terms of different types of Music Like Reggae, electronic classical etc ???

I like Ogg Vorbis too and I think that though it may produce artifacts here and there and probably a bit more often than other coders, I seem to find its artifacts more pleasing than those that accompany mp3, aac, and even mpc (at low bitrates). mpc at lowish bitrates sounds very harsh while Ogg Vorbis sounds nicer and the artifacts are less annoying.

In terms of the effect, there are two ways of looking at it: theoretical and subjective. It would seem in my mind to be a pointless exercise to talk solely on the theoretical effects since it will be our ears that will be the best judge. I'm sure the audio experts here can tell you what types of music MDCT-based transform coders work best and worst on as they have more experience. In other words, I dont know the answer to your question so I too am curious