Hi,

I was wondering whether anyone can tell me about the quantisation of LSPs in the G.729 speech coder. I don't have access to the ITU-T specs and have read in bits and pieces that it is a split multistage vector quantiser (S-MSVQ). Is this S-MSVQ of the same form as that used in the AMR-WB speech codec, where it is essentially multistage VQ with split VQs in each stage? Or is it a partial version, where only the second stage is split?

Many thanks,

QK

According to the name, it sounds like it's similar to what Speex is doing:
- pass1 : VQ codebook of dimension 10 for all LSP coeffs
- pass2 : 2 VQ code books (each of dimension 5) for the lower/upper coeffs
(In the second pass the errors of the first pass are quantized and coded.)

At least this is what I think Speex is doing. I don't know for sure if it's called S-MSVQ.


SebastianG