[Speex-dev] Comparison

[Jean-Marc Valin]

Le vendredi 10 juin 2005 à 21:27 -0400, SteveK a écrit :
> I'm not an expert either, but I see people choosing iLBC over speex  
> all the time with asterisk;  partly it's because they have more  
> market share in hardphones, and partly it's because of marketing and  
> such.  (another reason is that iLBC source is included in asterisk,  
> and speex is only compiled in if you have the speex development stuff  
> on your machine when you compile asterisk [i.e. the speex-devel  
> package]).   Some comments below:

Well, it's clear that the marketing department for iLBC is a bit larger
than that for Speex ;-) I can't comment on hardphone market share since
I have no idea about the iLBC market share (or even that of Speex).
However, I think that the (relatively recent) fixed-point should make
Speex much more interesting for embedded devices. 

> There's a wideband version of iLBC, but it's not free(as in beer).   
> Skype uses this.  (Actually I'm not 100% sure it's called speex, but  
> GIPS makes it, and it's probably closely related).

"Not sure it's called Speex"?? I suppose you meant something else?

> Even from their
> > site, you see that even at 13.3 or 15 kbps (they don't specify which),
> > they have the same quality as G.729A at 8 kbps.
> 
> The two different bitrates are just the difference between 30ms and  
> 20ms frame sizes; you use a bit more bps with the 20ms frames.

I'm aware of that, it's just that the quality may not be exactly the
same and they don't specify which bit-rate they use when comparing to
G.729A. In any way, I find it odd that they compare themselves to a
codec that has such a large difference in bit-rate.

> >
> > Now, their point is that since each frame is independent, the codec is
> > more robust to packet losses which is good for VoIP. However, in my
> > opinion (and that of others) is that the price in terms of bit-rate is
> > too high and you might as well just add redundancy by transmitting  
> > some
> > packets more than once (as proposed in
> > http://www.icassp2004.com/Papers/viewpapers.asp?papernum=3280 ). For
> > instance, if it goes to 20% packet loss, then transmitting 8 kbps  
> > Speex
> > packets twice (4% effective loss) will be much better than iLBC at 15
> > kbps.
> 
> I think you'd have to look at how Speex (using PLC) compares to iLBC  
> (using PLC) to make an accurate comparison here..

While I haven't compared with Speex, some people
http://www.icassp2004.com/Papers/viewpapers.asp?papernum=3280 have found
that if you take G.729A and start sending redundant information so that
you have the same effective bit-rate as iLBC, then all of a sudden
G.729A is *more* robust to packet loss. There's nothing magical about
it, there's just so much less loss. This is also despite the fact that
G.729A is not optimal for that task (it has a frame size of 20 ms, so
they lose several frames in a row when losing a 20 or 30ms packet). I
haven't done any formal testing, but I've listened to iLBC samples with
30% packet loss and Speex 8 kbps definitely sounds better than that at
9% loss (which is what you get when transmitting twice).

> > Regarding CPU utilization, I have no idea what iLBC requires. For a  
> > PC I
> > don't think it actually matters and for embedded systems, then you  
> > have
> > to pay for a fixed-point iLBC license (Speex includes the fixed- 
> > point in
> > the same code).
> 
> since there's so many speex options, it depends on the options you  
> choose.  Asterisk presently has poor defaults, and most people link  
> against speex-1.0 code.  If you use the latest speex code, and use  
> SSE optimizations, and complexity 2, you get very comparably  
> performance from speex as compared to iLBC.  The iLBC reference code  
> is not well optimized, though.

Not sure how speed is that relevant on a PC. With the current version,
you can encode+decode between 50 and 150 channels (depending on
quality/complexity), which is probably enough for most applications.

	Jean-Marc

-- 
Jean-Marc Valin <Jean-Marc.Valin at USherbrooke.ca>
Université de Sherbrooke