Lossy Audio Compression => Speech Codecs => Topic started by: audio_geek on 2006-02-09 04:31:15

Title: CELP a waveform coder... why ?
Post by: audio_geek on 2006-02-09 04:31:15
Hi, can you please tell me why CELP is a waveform coder ??
you could discuss the facts behind CELP being a waveform coder and/or suggest me links tutorials to quickly understand this fact....

Title: CELP a waveform coder... why ?
Post by: SebastianG on 2006-02-09 09:30:02
The key to that answer would probably be to think about the definition of "waveform decoder" and how this applies to the inner workings of CELP.

I'm not familar with that term (waveform coder) but if you look at this page http://www-mobile.ecs.soton.ac.uk/speech_c...s/waveform.html (http://www-mobile.ecs.soton.ac.uk/speech_codecs/waveform.html) you get an idea.

My understanding is this: waveform coder = opposite of parametric coder. Thus, PCM, MP3, Vorbis, CELP are all waveform coders and things like LPC10, MPEG4 HVXC/HLIN are parametric coders.

The thing that makes them different is: MP3(*), Vorbis(*) and CELP transform the signal into another representation WITHOUT loss and then quantize the outcome whereas parametric coders transform LOSSY the signal into a very short (parameter-)representation that describe an instance of a certain model and then quantize the parameters. Quantization is always lossy, but the transform may be lossy or lossless.

Although speech codecs like CELP tend to perform better on speech they can represent any signal if the bitrate is high enough in theory. The "transform" is specialized for speech though (in this case transform being linear prediction coding + long term prediction).

(* exceptions apply: intensity stereo would be a lossy transform but this is optional)

Title: CELP a waveform coder... why ?
Post by: itv on 2006-02-09 11:09:37
CELP is actually a hybrid codec, a mix between waveform and source codec. It models the vocal tract as source codecs do, but the exitation signal to that vocal tract model is selected so that the final resulting signal is as close to original as possible. Click few pages forward from the link Sebi provided to see more detailed description.