Skip to main content

Notice

If you are using a Hotmail or Outlook email address, please change it now, as Microsoft is rejecting all email from our service outright.
Topic: Vorbis Technology Demonstration (Read 32371 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Vorbis Technology Demonstration

I had a little fun with the Vorbis 1.0 libs, hacked a bit with it, and made some demonstration oggs of what it is capable of at the really low end.

Have a listen to:

http://sjeng.org/ftp/vorbis/scar.ogg  (6.4kbps)
http://sjeng.org/ftp/vorbis/queen.ogg (5.3kbps)
http://sjeng.org/ftp/vorbis/ss.ogg  (6.3kbps)
http://sjeng.org/ftp/vorbis/waitress.ogg (6.7kbps)

Edit: also

http://sjeng.org/ftp/vorbis/daan.ogg (6.6kbps)

And keep in mind this is 1/20'th of an 128kbps file.

--
GCP

Vorbis Technology Demonstration

Reply #1
:jawdrop:
Holy f'n $#!+!!!!  That's incredible!

Don't get me wrong, they sound like ass.  They sound remarkably good for 8kb/s though.

How much encoder tuning have you done, and is more possible?

Vorbis Technology Demonstration

Reply #2
Quote
Originally posted by Ardax
They sound remarkably good for 8kb/s though.
How much encoder tuning have you done, and is more possible?


6kbps, not 8kbps

There is no real tuning in here; I spent about an hour re-familiarizing myself with the 1.0 infrastructure and made the new encoding mode on guesswork. Getting everything to work took another two hours. I need to talk to monty to get a few things cleared up before more can be done.

I would like to put a CD on a floppy. If you have 2.8M floppies, you can do so already.

--
GCP

Vorbis Technology Demonstration

Reply #3
and make Vorbis Floppy-Transparent to 99% users over 99% samples:
"OGG Vorbis is the only format to give You floppy-quality™ at ~6kbps"

Vorbis Technology Demonstration

Reply #4
you can imagine what can be possible for 64 kbps :eek:

Vorbis Technology Demonstration

Reply #5
Impressive Garf indeed! I think most of the pure tones in Scar sound particularly very impressive at 8 kHz telephony fidelity. Have you tested this out with any sort of speech signal yet? You didn't happen to by any chance use floor 0 did you for testing purposes?
budding I.T professional

Vorbis Technology Demonstration

Reply #6
Setup an ogg stream (q0 40~45kb/s 22khz )

rexit2.ath.cx:8000/q0.ogg

(seems to be a GREAT alternative to mp3 streaming @ lower bitrates)

Vorbis Technology Demonstration

Reply #7
Quote
Originally posted by HotshotGG
Impressive Garf indeed! I think most of the pure tones in Scar sound particularly very impressive at 8 kHz telephony fidelity. Have you tested this out with any sort of speech signal yet? You didn't happen to by any chance use floor 0 did you for testing purposes?


I've tested it with speech against Speex 5.1.

The size comes around 5-6kbps, which is lower than Speex quality 1 (goes from 1-10) and higher than Speex quality 0. Speex quality 0 artifacts quite badly and I wouldn't know if this mode or Speex is better. At quality 1, Speex wins hands down. The problem of this mode is that it doesn't window-switch, which (I think) is needed to prevent sibilants from echoing. Except for the echo's, the mode sounds quite reasonable on voice.

Based on this result, I think you could make a Vorbis 1.0 encoder that handles voice at least as well as the current Speex. Interesting result, considering Speex was designed for voice and Vorbis wasn't. You'd want to have someone with more knowledge about voice coding than me working on it though

I would have liked to try floor0 with this, as it _possibly_ could be a little more efficient, but the 1.0 code doesn't have any other mode that uses it! Without an example or prior codebooks to work from it is impossible for me to make one. I already had enough problems with the floor1 codebooks now.

--
GCP

Vorbis Technology Demonstration

Reply #8
Quote
Without an example or prior codebooks to work from it is impossible for me to make one. I already had enough problems with the floor1 codebooks now.


yes, in deed I had been looking over the floor 1 backend implementation myself and I still need to look over it a few more times to figure it out. I wouldn't argue with you about the codebooks as they look they are a major task to figure out. I had wanted to point that out with Floor 0, because floor 0 as you know uses a IIR LSP representation. This might be especially useful in speech coding. It seems as though floor 0 is deprecated for the time being though like you said. I had been looking through the documentation for it though anyway.
budding I.T professional

Vorbis Technology Demonstration

Reply #9
Quote
Originally posted by HotshotGG
 

I had wanted to point that out with Floor 0, because floor 0 as you know uses a IIR LSP representation. This might be especially useful in speech coding. 


The usage is so different that success of LSP/LPC in speech coding has no bearing on Vorbis. I'd prefer floor0 because it may be more efficient for uncoupled modes (like mono) than floor1, even when not encoding speech.

--
GCP

Vorbis Technology Demonstration

Reply #10
Quote
The usage is so different that success of LSP/LPC in speech coding has no bearing on Vorbis. I'd prefer floor0 because it may be more efficient for uncoupled modes (like mono) than floor1, even when not encoding speech.


I see where you coming from. It might be useful from an experimental point. I had figured that LSP/LPC representation to have no bearing for speech coding in Vorbis. I had just been quite confused when it had been implemented and I was reading it was popular in speech coding. Now I understand everyone has there own flavors of it.
budding I.T professional

Vorbis Technology Demonstration

Reply #11
Uhm...could you give us the exact command line for these ultra-low settings? I'd like to do previews of entire albums for my band at cdstreet.com. We have 10 MB of space of samples of the CDs and with that bitrate (which sound damn good) I can fit a ton of stuff in there...

so, what q- setting is this???? -1? -2? -3?

Vorbis Technology Demonstration

Reply #12
OMG!  LOOK HOW SMALL AND CUTE THEY ARE!  6KBPS!!! WOW!  THEY ARE LIKE LITTLE BABIES!

Vorbis Technology Demonstration

Reply #13
Quote
Originally posted by Zaraza
Uhm...could you give us the exact command line for these ultra-low settings? I'd like to do previews of entire albums for my band at cdstreet.com. We have 10 MB of space of samples of the CDs and with that bitrate (which sound damn good) I can fit a ton of stuff in there...

so, what q- setting is this???? -1? -2? -3?


This is a custom version of the encoder libs. If I can get monty to help sort out a few remaining issues, I'll make it available.

--
GCP

Vorbis Technology Demonstration

Reply #14
Quote
Originally posted by Zaraza
We have 10 MB of space of samples of the CDs and with that bitrate (which sound damn good) I can fit a ton of stuff in there...


oggenc -q -1 --downmix --resample 8000

with the normal Vorbis 1.0 oggenc will already fit an entire CD in 10M

--
GCP

Vorbis Technology Demonstration

Reply #15
Quote
If I can get monty to help sort out a few remaining issues


What would you need sorted out might I ask?
budding I.T professional

Vorbis Technology Demonstration

Reply #16
Quote
Originally posted by HotshotGG
 
What would you need sorted out might I ask?


Why I get massively differing performance from one codebook training run to another.

Why some samples completely fall apart (as in: not remotely comparable to original) even though they're not particularly hard.

Why the codebook trainers sometimes don't produce all the residue training files I'd expect them to.

What the correct parameters for floor training are (I guessed them).

The codebook training part of Vorbis is one of the least intiutive (at least to me) and it's almost undocumented (the spec doesn't help either)

--
GCP

Vorbis Technology Demonstration

Reply #17
Quote
The codebook training part of Vorbis is one of the least intiutive (at least to me) and it's almost undocumented (the spec doesn't help either)


yes, when I was looking over the source code myself I didn't even make an effort to look over the codebooks as well I had the same basic idea in my mind. What codebooks are you using just the regular scalar huffman codebooks?
budding I.T professional

Vorbis Technology Demonstration

Reply #18
Quote
Originally posted by HotshotGG
 
What codebooks are you using just the regular scalar huffman codebooks?


IIRC, floor codebooks are basic huffman and residue is lattice VQ

--
GCP

Vorbis Technology Demonstration

Reply #19
Quote
IIRC, floor codebooks are basic huffman and residue is lattice VQ


Thank you for bringing that to my attention again. I was aware of the residue using lattice VQ codebooks, however I never really thought that the floor codebooks used huffman coding. I see the puzzle fits together now.
budding I.T professional

Vorbis Technology Demonstration

Reply #20
I was fooling around with this extreme low bitrate stuff last week. Mono, -q -1, 8KHz resampling, etc. I encoded George Winston's December CD, which is all solo piano. Even at 8kbps, it didn't sound half bad. I guess you could use this to stream music over a 28.8kbps modem connection...but who has those anymore?

<- 1.5Mbps broadband pipe

Vorbis Technology Demonstration

Reply #21
Quote
which is all solo piano


yes, strong tones sound particularly good it's unfortunate there is no window switching like Garf mentioned above to eliminate any echo's etc.


Quote
<- 1.5Mbps broadband pipe


My dad is heavenly into Broadband and from a growing response in that area I think we are moving in just the opposite direction. I can't even get my friend back on his 56.6 V. 90 dial-up. AT&T is ripping everyone off in this city because there is no competition, but that's another story that is more or less important for a discussion on this forum because physcoacoustics audio compression is more important.]
budding I.T professional


Vorbis Technology Demonstration

Reply #23
It's hard to believe ! Absolutely amazing (but really awful, too) 
Can you go below ?

Vorbis Technology Demonstration

Reply #24
Amazing. Though there are a few codecs designed for voice transmission in the 1000-6000bps range, I don't recall any of them ever being able to handle a musical signal this well when playing around with them (voxware, 1-bit ADPCM, etc). An application for such a low bitrate could include streaming audio for low-bandwidth digital cell phones (9-14.4kbps). Is it possible to get an 8Khz stereo encode at bitrates under 10kbps?

If this demonstration were run by Micro$oft, they would probably claim "CD Quality  At 3kbps!" 

 
SimplePortal 1.0.0 RC1 © 2008-2021