[vorbis-dev] Vocoders

[Steve Underwood]

LPC-10 was hardly the last effort. Many codecs have followed that one.

The original full rate GSM has the big benefit of being widely deployed and free
from IP troubles (it has to be - its so old any patents would have run out by
now). Its bit rate is a bit high, though, and its quality isn't too impressive.
Its only substantial plus point for quality is that almost any coder at a lower
rate totally falls apart when coding anything but a single human voice. Full rate
GSM only half falls apart.

Half rate GSM, at around 6kbps, is far more interesting. Reasonable bit rate,
excellent quality, and masses of IP encumberance. I know (I did some work on it)
that is scales well down to around 4kbps, too. Bit high on CPU grunt, but nice
results.

G723.1 isn't too bad, and doesn't take so much CPU grunt, but still has just as
much encumberance.

The Qualcomm codecs used in US narrow band CDMA are not too bad, except for the
IP issues.

MCELP (which TI used in some digital answering machine chips, and which I think
is basically similar to the US government MELP codec) isn't too good at 4kbps, so
it might be interesting to hear what the MELP codec is like at 2.4kbps. I'm
puzzled that anything as low as 2.4kbps would be a form of ....ELP, as it takes a
large percentage of that bit rate just to code the VQ'd LP coeffs. Anything much
lower than 4kbps is pushing it if you want to maximize quality. The best things
I've heard at those rates have been based on some form of VQ'd sub-banding
tailored to a model of the human vocal tract (rather than being a direct physical
model of the human vocal tract, like ....ELPs). Anyway, I'll have to play with
this code and see what it can do out of interest. Its of little other value, due
to good old bad old IP issues.

Ah..... the search for an unencumbered, high quality, 4 to 6 kbps codec
continues. I feel that starting such a project myself is a non-starter for 2 key
reasons. First is the obvious time required. Second is I can't think of any
sensible strategies which are in either the non-yet-patented or expired-patent
categories. Stinker, huh.

Does anyone know what techniques the low bit rate vocoder in the
MPEG-something-or-other-recent-issue uses, and what encumberance there may be? I
guess if MPEG is getting a low bit rate vocoder mode, Vorbis will need one
too......................

Steve

seanl at literati.org wrote:

> The DoD has a 2400 bps vocoder with quality similar to 4kbps CELP.  It's
> called Mixed Excitation Linear Predictor.  I've played around with it and
> it has excellent quality (though nowhere near GSM; you can still tell
> you're using a vocoder) and handles background noise pretty well.  This is
> definitely more recent than LPC-10.  GSM is 13 kbps, so I suppose it
> depends on what sort of bitrate and quality you want.
>
> http://www.plh.af.mil/ddvpc/index.html
>
> This page has MELP, LPC, CELP, and CVSD.  I didn't see source code there,
> so if you want source code for MELP, email me and I'll track it down (I
> have it on a machine without Internet access right now).
>
> On Wed, May 10, 2000 at 05:27:27PM -0700, Rage O Matic wrote:
> > > > Vorbis looks real interesting. but I'm more interested in voice coders
> > > > right now. I worked on some really nice stuff at around 4kbps in the
> > > > early 90's. The trouble is any coder even remotely similar is surrounded
> > > > by an impenetrable shield of IP. Does anyone know of any work in
> > > > progress on good free low bit rate vocoders?
> > >
> >
> > I dont think any low bit rate vocoders are in progress now. LPC-10 was the
> > last effort I've seen. Not much work on those algorithms. GSM seems to be
> > the most recent, and most advanced coding. It's used int most cell phones.
> > ADPCM is another group of compressions.

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/