Experimental release of Ghost/CELT 0.0.1 Options

[jmvalin]

I've just made the first public release of some new *experimental* codec work I've been doing (part of the vague Ghost project) with help from Monty and Timothy. This is mainly intended for developers with DSP knowledge, not for doing anything useful with it (but it does encode and decode already). Also, the main idea is *not* to replace either Speex or Vorbis, but to code audio with really low latency -- currently 8 ms.

This is still very experimental and everything is still likely to change, including the exact goals. The algorithm is called (temporary name) Code-Excited Lapped Transform (CELT) and the main ideas are:
- Using an MDCT on very short frames
- Dividing into 15 bands and transmitting the energy for each band
- Using a pitch predictor (good for speech, but helps for music as well).
- The rest is coded using a unit-pulse codebook.
At this point, I'm still trying to figure out how to fit psychoacoustics into this.

CELT is based on a paper I submitted to ICASSP and which I'm hoping will be accepted so I can make it available to everyone. The only difference is that the ICASSP paper was based on the FFT (non critically sampled), whereas this version is based on the MDCT. One part that is already published though is Tim's explanation of the pulse codebook encoding.

The full source for CELT is available at: http://downloads.us.xiph.org/releases/celt/celt-0.0.1.tar.gz or through git at http://git.xiph.org/celt.git

I've put some music samples at 56 kbps CBR at http://people.xiph.org/~jm/comp_celt58cbr.wav with the original at http://people.xiph.org/~jm/comp44.wav . As you can hear, it definitely doesn't suck as much as Speex on music, but there's still room for improvement.

I'm open to interesting ideas, but don't bother complaining if it doesn't work or if it explodes in your face :-) Oh, and don't expect a final codec any time soon.

Have fun!

[radorn]

So this codec is not intended to eventually replace vorbis?

From what monty said here, I thought that was the case.

This post has been edited by radorn: Dec 20 2007, 14:42

[jmvalin]

So this codec is not intended to eventually replace vorbis?

From what monty said here, I thought that was the case.

What I mean is that in its current form, there's no way it'll beat Vorbis. However, if you add sinusoidal coding and use CELT just to encode the noise, that's another matter. That's what Monty is currently hoping to do with Ghost. Whether we'll end up with one codec or two is still an open question. Even then, a codec that will replace Vorbis is still a long way off, so don't hold your breath.

For those interested, I've posted this overview of CELT along with some technical details on energy coding.

[SebastianG]

For those interested, I've posted this overview of CELT along with some technical details on energy coding.

Nice!

What's the rationale behind "pulse code books"? I read Timothy's pulse coding web site and I'm not sure of what he's trying to do. Is it an attempt to uniformly tesselate the unit sphere or only a code book specialized tor "pulsy" signals? Judjung by the way it is used in CELT I'd say it's the former -- which is weird because at first glance the formulas don't look like the resulting codebooks do tesselate the unit sphere well. It might be me as I'm not familiar with those kind of specialized vector code books.

Regarding pitch prediction: Am I correct when I say what you do is pretty much the same thing as LTP (MPEG4)?

Cheers!
SG

[jmvalin]

What's the rationale behind "pulse code books"? I read Timothy's pulse coding web site and I'm not sure of what he's trying to do. Is it an attempt to uniformly tesselate the unit sphere or only a code book specialized tor "pulsy" signals? Judjung by the way it is used in CELT I'd say it's the former -- which is weird because at first glance the formulas don't look like the resulting codebooks do tesselate the unit sphere well. It might be me as I'm not familiar with those kind of specialized vector code books.

In the CELT context, the pulse codebook has the following properties:

• Decent (though suboptimal) tesselation of the sphere
• algebraic representation (no storage)
• slight "bias" towards impulses (i.e. tonal signals)
• Fast recursive search even for huge codebooks (one pulse at a time)

Of course, none of that is frozen, so I'm always open to other ideas. Did you have something in particular in mind?

Regarding pitch prediction: Am I correct when I say what you do is pretty much the same thing as LTP (MPEG4)?

I'm not aware of any other codec doing things like I'm doing, but then again, I've no idea what this mpeg4 LTP is or does. Can you give some details about that codec? BTW, the next part on CELT will be on the pitch prediction.

[SebastianG]

In the CELT context, the pulse codebook has the following properties:
(...)

Oh, okay. I almost forgot the "algebraic representation" thing. You're right, this is a big advantage.

Of course, none of that is frozen, so I'm always open to other ideas. Did you have something in particular in mind?

No, sorry. None that fits the strict gain/shape separation you are striving for. But there is something that I always wanted to see tested: A combination of trellis-coded quantization with subtractive dithering. "Subtractive dithering" is achieved by a randomized trellis graph using the same pseudo random number generator within both, the encoder and decoder. I believe it's worth a try because

• En- and decoding is fast and doesn't require big tables either
• It's possible to fade seamlessly between PNS (perceptual noise substitution) and high rate coding.
• No odd (ie metallic sounding) artefacts besides white noise even on large MDCT blocks (dithering is "the right thing to do"^TM which unfortunately got lost on the way to the lossy coding world -- until now -- *hint hint*) ;-)
• it should be possible to keep the quantized signal's energy close to the original's even at very low rates (thanks to subtractive dithering and adaptive rate-distortion optimization)

I'm not aware of any other codec doing things like I'm doing, but then again, I've no idea what this mpeg4 LTP is or does. Can you give some details about that codec? BTW, the next part on CELT will be on the pitch prediction.

LTP is an MPEG4-AAC tool and is short for long term prediction. The decoder just remembers the past decoded samples. The encoder might signal that the following block looks similar to one that was xyz samples ago (aka "pitch period"). Then the decoder would do an (forward) MDCT on the block referenced by the "pitch period", apply frequency-adaptive prediction gains and subtract this from the current MDCT block's samples.

I'm not sure right now but I guess LTP can be coupled with LD-AAC (low delay, 480 samples/frame) which would then be very similar to your CELT except for the special pulse codebook.

Cheers!
SG

This post has been edited by SebastianG: Dec 23 2007, 12:24