Wavelets, why use in vorbis? Options

[MGuti]

as i have heard, wavelets are to be implemented at some point in vorbis (if ever). I did some reading about them (most of it i didn't understand) and was curious as to how it would be used to encode music.

[QuantumKnot]

Wavelets have certain desirable properties that make them attractive in coding. First thing is that they can be easily implemented in a subband coding algorithm (like in MPC or MPEG-1 Layer 2) so they enjoy low complexity decoding. Secondly, they have a good balance between temporal resolution as well as frequency, so they tend to handle transients better than transform coders, which have great frequency resolution but terrible temporal resolution.

This post has been edited by QuantumKnot: Feb 11 2004, 01:34

[MGuti]

are they difficult to implement? have they been implemented in any other codecs?

[QuantumKnot]

are they difficult to implement? have they been implemented in any other codecs?

Not at all. Implementing wavelets is often as simple as replacing the filter coefficients of a subband coder.

I'm not sure if wavelets have been used in other audio coders (I assume MPC and MPEG-1 Layer 2 use QMFs) but I'm sure they've been tried before since they were heavily hyped in the signal processing community over the past few years. Wavelets were tried in video coding, image coding, speech recognition, etc. The most successful application of wavelets has been in image coding where they are astronomically superior to transform coders. I don't think they has found a place in audio (yet).

I guess the good thing about Vorbis II is, when Monty does get the wavelets to work, the adoption of it will be easier since Vorbis I has already made a moderate impact. Essentially you could say, Vorbis I has provided the customer base for Vorbis II, a wavelet-based coder, which is certainly an advantage over some unknown company that has released their own unknown wavelet coder.

[rjamorim]

are they difficult to implement? have they been implemented in any other codecs?

http://www.hydrogenaudio.org/forums/index....indpost&p=27546

[Doctor]

I distinctly remember somebody saying, perhaps even on this board, that attempts to implement wavelet-based audio compression did not provide any significant advantage over other methods.

Edit: thanks Roberto, that's what I meant. ;-)

This post has been edited by Doctor: Feb 11 2004, 02:20

[QuantumKnot]

I distinctly remember somebody saying, perhaps even on this board, that attempts to implement wavelet-based audio compression did not provide any significant advantage over other methods.

Edit: thanks Roberto, that's what I meant. ;-)

Or perhaps you meant the thread about the telephone conversation with Frank Klemm where he said he sees no advantages of subband coders (eg. MPC) over transform coders (eg. AAC, Vorbis), other than the psychoacoustic model.

This post has been edited by QuantumKnot: Feb 11 2004, 02:24

[rjamorim]

Or perhaps you meant the thread about the telephone conversation with Frank Klemm where he said he sees no advantages of subband coders (eg. MPC) over transform coders (eg. AAC, Vorbis), other than the psychoacoustic model.

Nope. He was probably referring to this post:
http://www.hydrogenaudio.org/forums/index....indpost&p=26677

And, most importantly,
http://www.hydrogenaudio.org/forums/index....indpost&p=24193

This post has been edited by rjamorim: Feb 11 2004, 02:59

[Doctor]

Roberto: exactly. ;-)

[MGuti]

so, from what i've read here, theres no real reason to get excited about wavelets in vorbis. although it could potentially improve quality in some areas, it is no better than the commonly used methods.

alas, and i was excited for a while about what wavelets might have accomplished for vorbis (if it was so great everyone would be using it i guess).

[rjamorim]

(if it was so great everyone would be using it i guess).

Very good point. MPEG seems to be rather fast to add state-of-the-art technologies to their codecs (SBR, PS, H264, etc.)

This post has been edited by rjamorim: Feb 11 2004, 03:26

[QuantumKnot]

Compared with images, audio certainly is a different beast.

[eltoder]

At least wavelets have advantage over PQF, because they're lossless and hence do not add noise. Also, you can make non-uniform transform with different time-frequency resolution in different parts of the spectrum (some static choices like one band for 11025-22050 to use some sbr-like method, or dynamic choices - better time resolution in parts with transients and better frequency resolution in tonal parts - something like mp3's mixed blocks, but easier to implement). But since audio signals usually have a lot of separated frequencies wavelets do not have such great advantages like they do in image coding.

-Eugene

[QuantumKnot]

Just a minor sidepoint....the new buzzword in image coding literature these days seems to be 'curvelets'. I'm sure it will be a very long time till we see much mention of 'curvelets' for audio coding.

[Ivan Dimkovic]

Despite some theoretical backings that Wavelets might be better than subband/transform filterbanks in audio codecs (namely because of possibility of designing T/F resolution similar to critical band analysis in the human ear) so far there has not been >any< proof that a wavelet audio coder might achieve better performance than current state of the art.

Except some marketing claims from various industry start-ups (that are probably gone now), and few university projects - I haven't seen any successul solution that could suprass even MP3, leave alone some better coding solutions.

In the last scientific test I know of, it has been shown that switched 128/1024 MDCT filterbank has best properties for audio coding (this filterbank is used in AAC and Vorbis - and probably WMA)

I'd like to see some breakthrough there, but I am not very optimistic about wavelet technology as a filterbank in audio coder.