dear all !!!


I have the specification ITU g.728 , I have to implement this codec for my work but i am having the problem while geting the flow of the G.728 encoder this algo is based on the LD-CELP theory. First I want to know what is CELP

wnd what is the diffrence between the CELP and the LD - CELP and why it came into exsistance.

And what is the basic knowledge ( flow of the G.728 codec )

please guide .. as i am new to this field..

ranjeet

No need to post the same question in two different places...

CELP - Code Excited Linear Predictions. Very similar to GSM as it is linear prediction based.. However the residual is model as a codebook.

LD-CELP - I am not very sure of what LD means.. Maybe Low Delay? If so, then as far as I know, there are 2 ways of conducting linear predictions..

Foward Predictions and Backward Predictions. In Foward Predictions, the LPC / LSP coefficients must be transmitted to the decoder. In Backward Predictions, no coefficients are transmitted resulting in coding gain. There is a adaptation algorithm at the decoder will estimate the current audio samples based on previous reconstructed samples..

I think I have read something about backward predictions in speech coding.. but backward predictions can be highly unstable compared to foward predictions.. Anyway I cannot be sure.

As someone pointed out, CELP is "code excited linear prediction" and LD-CELP is "low-delay" CELP. Actually, I'd say LD-CELP is closer to ADPCM than CELP because everything is backward-adaptive (the main difference being that instead of coding samples one at a time, they're coded with 5-sample subvectors.

Now, if you are new to the field, I doubt very much you'll be able to implement the code with just the specs (no reference code). There are just so many places where you can screw up things. Believe me, I screwed up so many times when writing Speex -- and I wasn't even trying to match a spec exactly.

In foward prediction based LPC coders, there is a need to "frame" the speech signals for processing.. at both the encoder and decoder.. This results in an algorithm delay of between 20 - 30 miliseconds.. the time lag between the transmitted speech and the listener.. (assuming > realtime processing..) This delay can be very annoying in real-time 2 way conversations.. As a result, the quest has been to minimized this delay as much as possible for better 2 way realtime conversations over telecommunications networks.. That is where Backward Predictions comes in..

Well, I have worked with MPEG Committee experts before.. and I was told that the specifications includes information from the "ISO conformance decoder reference source codes".. In other words.. much of the needed information to construct an encoder must be "reversed engineered" from the ISO decoder source codes..

The ISO specifications isn't enough..