[Speex-dev] Re: Speex, ACELP, G.729

[Jean-Marc Valin]

> I've been watching the speex development from its inception because I 
> and several Tech Startup Connection members have a very important 
> application for this voice encoding/decoding. Further, we are quite 
> familiar with ACELP as implemented in G.729. As far as I know Speex 
> is also ACELP ... yes/no?

No. Speex is CELP (Code-excited linear prediction), but not ACELP
(Algebraic CELP). I have nothing against ACELP, but I couldn't use it
because of patent issues.

> Question ... what do you see as the advantages of ACELP compared to 
> MP3? I had thought that ACELP yielded much smaller files, but now I'm 
> not sure because I've listened to what I think were 16kbps MP3 files 
> (????) with Apple's ITunes that sounded pretty good ... comparable to 
> G.729's 8kbps. I'm speaking about G.729 because I'm not up on Speex 
> attributes ... e.g. what Speex bps compares to G.729's 8kbps quality.

I think (my opinion) Speex 11 kbps is a bit better than G.729, while
Speex 8 kbps is a bit worse. Speex 8 kbps VBR is probably equivalent. As
for 16 kbps, I think it's terrible, you probably had AAC, not MP3 in
your 16 kbps test. Another thing to keep in mind is that there is more
to coding than file size. Speex (like other speech codecs like G.729) is
designed for real-time communications. The frame size is 20 ms, so it
has a much lower delay than MP3. Also, the Speex encoder can run in
real-time with little complexity (e.g. it runs real-time on ARM). The
last thing is that Speex is designed to handle packet losses, something
that I don't think has been done to MP3.

	Jean-Marc

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