Vorbis dev - Jan 2002 - Just to dispel any hopes -- RC3 really low bitrate

I've just done some rudimentary testing to see how Vorbis degrades at
absurdly low bitrates without downsampling. In summary, don't hope for
anything decent below -q 0 for now. I tried oggenc -b <bitrate> -M
<bitrate> for the below and a few in between:

24k - spectral energy "floor" captured decently, but many pure-tone
blips (think old computer movie sound effects) in mid to high
range. Totally untolerable.

...

35k - blips mostly gone, audio present but very muffled (high
frequency basically gone). Occasionally it will "open up" a little and
get some high-end, and stay like that for maybe a second (definately
more than one block), then close up again. I might be able to live
with this, it sort of sounds like Real at 16k or so.

...

45k - now most of the time it sounds like 35k when it opened up, and
like 35k there are times when it'll pull in another block of
higher frequencies. The average quality level seems about consistant
with what I'd expect a stream coming down from my 28.8 modem (~24k) to
sound like, if the rough edges were smoothed over a little.

...

64k - just to try what (ideal) single-channel ISDN would sound like;
generally quite tolerable. Still munges high frequency, but nearly an
order of magnitude less perceptably than lower bitrates. Note that
this is not the VBR 64k mode; I've essentially locked the codec at 64k
and below. It still should sound roughly like 64k normally (which I
can live with for e.g. background music), but it has more obviously
bad spots.

Definately some improvement in tuning elsewhere too. Quality around 2
gets me roughly the same result I've come to expect from 128k --
pretty much transparent unles listening very closely -- but at around
30kb/s less than RC2. Those with more "golden ears" than I may contest
that as a general statement, but I think I'll encode at around 2.5 for
most things.

For future Vorbis development, I think the quality scale should be
shifted lower, or perhaps made logrithmic -- the marginal benefit per
unit bitrate is much more at lower quality settings than higher (I
doubt I would be able to hear any difference at all for 5 and above,
while 0 through 3 were obvious step-ups). Also, if possible lower
bitrates should degrade more gracefully. Gradually add
high-frequencies, not in chunks. Does the bitrate limiting engine have
any linkup with psycoacoustics? -- it doesn't sound like it, at least
not for lower bitrates. I'd experiment with adding in harmonics of the
highest frequency coded to fill in the high end; it may be better to
have a wrong high-end than no high-end, psycoacoustically speaking.

Note that the Vorbis internals have changed significantly since I
last really tried to study them, so pretend that I don't know what I'm
talking about and you'll probably be right.


-- 
Kenneth Arnold <ken@arnoldnet.net>
- "Know thyself."


-------------- next part --------------
A non-text attachment was scrubbed...
Name: part
Type: application/pgp-signature
Size: 233 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/vorbis-dev/attachments/20020101/873399bb/part-0001.pgp

On Tue, Jan 01, 2002 at 10:25:25PM -0500, Kenneth Arnold
wrote:
> I've just done some rudimentary testing to see how Vorbis degrades at
> absurdly low bitrates without downsampling. In summary, don't hope for
> anything decent below -q 0 for now. I tried oggenc -b <bitrate> -M
> <bitrate> for the below and a few in between:
That's not quite the best way to do it; just -M with no -b will give
somewhat better results.  In rc4, -q with -M will be possible, and
this will be better yet.  Setting -b and -M to the same value will
cause far too much drastic hard limiting of the audio.  You want the
codec to be able to 'work ahead' and save up bits in the easy
sections; using -b says 'always use 'em'.

In any case, it's possible to go under 64kbps 44.1kHz stereo, but I
find the results harsh so the encoder is neither tuned for it, nor
intended to go lower.  I can provide some examples using the rc4
branch if you're curious; the effect is actually mp3proish as the high
end gets more and more synthetic sounding, as opposed to the near
total falling apart it does now.

Monty

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.

in case your curious as to what low bitrates sound *with* resampling, check 
out http://www.oddsock.org/vorbis_streams

there are 2 streams there, one with RC3 32/22/Mono, and one with LAME 
32/22/Mono....the source is the same for both streams...I personally think 
that the 32/22/Mono of RC3 is quite spectacular...

oddsock
At 10:25 PM 1/1/2002 -0500, you wrote:
>I've just done some rudimentary testing to see how Vorbis degrades at
>absurdly low bitrates without downsampling. In summary, don't hope for
>anything decent below -q 0 for now. I tried oggenc -b <bitrate> -M
><bitrate> for the below and a few in between:
>
>24k - spectral energy "floor" captured decently, but many
pure-tone
>blips (think old computer movie sound effects) in mid to high
>range. Totally untolerable.
>
>...
>
>35k - blips mostly gone, audio present but very muffled (high
>frequency basically gone). Occasionally it will "open up" a little
and
>get some high-end, and stay like that for maybe a second (definately
>more than one block), then close up again. I might be able to live
>with this, it sort of sounds like Real at 16k or so.
>
>...
>
>45k - now most of the time it sounds like 35k when it opened up, and
>like 35k there are times when it'll pull in another block of
>higher frequencies. The average quality level seems about consistant
>with what I'd expect a stream coming down from my 28.8 modem (~24k) to
>sound like, if the rough edges were smoothed over a little.
>
>...
>
>64k - just to try what (ideal) single-channel ISDN would sound like;
>generally quite tolerable. Still munges high frequency, but nearly an
>order of magnitude less perceptably than lower bitrates. Note that
>this is not the VBR 64k mode; I've essentially locked the codec at 64k
>and below. It still should sound roughly like 64k normally (which I
>can live with for e.g. background music), but it has more obviously
>bad spots.
>
>Definately some improvement in tuning elsewhere too. Quality around 2
>gets me roughly the same result I've come to expect from 128k --
>pretty much transparent unles listening very closely -- but at around
>30kb/s less than RC2. Those with more "golden ears" than I may
contest
>that as a general statement, but I think I'll encode at around 2.5 for
>most things.
>
>For future Vorbis development, I think the quality scale should be
>shifted lower, or perhaps made logrithmic -- the marginal benefit per
>unit bitrate is much more at lower quality settings than higher (I
>doubt I would be able to hear any difference at all for 5 and above,
>while 0 through 3 were obvious step-ups). Also, if possible lower
>bitrates should degrade more gracefully. Gradually add
>high-frequencies, not in chunks. Does the bitrate limiting engine have
>any linkup with psycoacoustics? -- it doesn't sound like it, at least
>not for lower bitrates. I'd experiment with adding in harmonics of the
>highest frequency coded to fill in the high end; it may be better to
>have a wrong high-end than no high-end, psycoacoustically speaking.
>
>Note that the Vorbis internals have changed significantly since I
>last really tried to study them, so pretend that I don't know what
I'm
>talking about and you'll probably be right.
>
>--
>Kenneth Arnold <ken@arnoldnet.net>
>- "Know thyself."
<p><p>--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to
'vorbis-dev-request@xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is
needed.
Unsubscribe messages sent to the list will be ignored/filtered.