hi all, i am puzzled about the use of long-time predictor in celp. i see in some implementations, there is a long-time predictor (ltp) but in others, eg fs-1016, there isn't. if I understand correctly, the ltp is used to compute the gain. but why is there a need to do this computation when eventually, in the codebook search, the gain and period of the pitch is "computed". thanks in advance. rgds, tk |
pitch predictor in celp
Started by ●February 21, 2005
Reply by ●February 23, 20052005-02-23
Your question is not quite clear. Long-term predictor is actually an adaptive codebook. In analysis-by-synthesis scheme the search in the codebooks (adaptive and stochastic) gives both the codevector and the scaling factor (gain). I guess that by LTP you mean something like Harmonic Noise Filtering (in g723), but it is not LTP. Ilya Druker --- In , tunkeat <tunkeat@g...> wrote: > > hi all, > > i am puzzled about the use of long-time predictor in celp. > i see in some implementations, there is a long-time predictor (ltp) > but in others, eg fs-1016, there isn't. if I understand correctly, > the ltp is used to compute the gain. but why is there a need > to do this computation when eventually, in the codebook search, > the gain and period of the pitch is "computed". thanks in advance. > > rgds, > tk |
Reply by ●February 23, 20052005-02-23
hi IIya, thanks for reply. in celp literature, sometimes I see a formant predictor (linear prediction) and a pitch predictor (long-term predictor), in others, there isn't one, eg in fs1016, they only have adaptive codebook for pitch prediction (to get gain and pitch period). in one paper from http://www.tsp.ece.mcgill.ca (excellent website on speech coding), it was mentioned that in some celp implementations, a 3-tap pitch predictor is commonlyused. however, it asserted that this is redundant as in the synthesis stage, the adaptive codeboook will be searched to predict pitch. thus, I am not sure what is the motivation behind using a LTP when we are already doing the search in the adaptive codebook. i can only guess for speeding up the search hope this is clearer. rgds, tk On Wed, 23 Feb 2005 05:32:22 -0000, Ilya Druker <> wrote: > > > Your question is not quite clear. Long-term predictor is actually an > adaptive codebook. In analysis-by-synthesis scheme the search in the > codebooks (adaptive and stochastic) gives both the codevector and the > scaling factor (gain). I guess that by LTP you mean something like > Harmonic Noise Filtering (in g723), but it is not LTP. > > Ilya Druker > --- In , tunkeat <tunkeat@g...> wrote: > > > > hi all, > > > > i am puzzled about the use of long-time predictor in celp. > > i see in some implementations, there is a long-time predictor (ltp) > > but in others, eg fs-1016, there isn't. if I understand correctly, > > the ltp is used to compute the gain. but why is there a need > > to do this computation when eventually, in the codebook search, > > the gain and period of the pitch is "computed". thanks in advance. > > > > rgds, > > tk > > To |
Reply by ●February 24, 20052005-02-24
Please note that LTP is not used to compute the gain. The short term predictor or STP removes short term redundancies where as the LTP removes the long term redundancies. You can say that the LTP is intended to and tends to capture the redundancies of pitch frequency which are longer term or roughly the periodicity you observe in the envelope of the signal. Regards Bala --- Ilya Druker <> wrote: > > > Your question is not quite clear. Long-term > predictor is actually an > adaptive codebook. In analysis-by-synthesis scheme > the search in the > codebooks (adaptive and stochastic) gives both the > codevector and the > scaling factor (gain). I guess that by LTP you mean > something like > Harmonic Noise Filtering (in g723), but it is not > LTP. > > Ilya Druker > > --- In , tunkeat > <tunkeat@g...> wrote: > > > > hi all, > > > > i am puzzled about the use of long-time predictor > in celp. > > i see in some implementations, there is a > long-time predictor (ltp) > > but in others, eg fs-1016, there isn't. if I > understand correctly, > > the ltp is used to compute the gain. but why is > there a need > > to do this computation when eventually, in the > codebook search, > > the gain and period of the pitch is "computed". > thanks in advance. > > > > rgds, > > tk > > __________________________________ |
Reply by ●February 24, 20052005-02-24
hi Bala, i am thinking aloud here as i am not sure. speech-> STP -> LTP-> residual (near gaussian) LTP sometimes called pitch predictor. for pitch prediction, normally 2 things are determined for the pitch filter. the pitch gain and the pitch period. when i say LTP computes gain, i mean the gain for this filter which is the pitch gain. is that right? rgds, tk On Thu, 24 Feb 2005 06:57:54 -0800 (PST), V. Balakrishnan <> wrote: > > Please note that LTP is not used to compute the gain. > The short term predictor or STP removes short term > redundancies where as the LTP removes the long term > redundancies. You can say that the LTP is intended to > and tends to capture the redundancies of pitch > frequency which are longer term or roughly the > periodicity you observe in the envelope of the signal. > Regards > Bala > > --- Ilya Druker <> wrote: > > > > > > > Your question is not quite clear. Long-term > > predictor is actually an > > adaptive codebook. In analysis-by-synthesis scheme > > the search in the > > codebooks (adaptive and stochastic) gives both the > > codevector and the > > scaling factor (gain). I guess that by LTP you mean > > something like > > Harmonic Noise Filtering (in g723), but it is not > > LTP. > > > > Ilya Druker > > > > --- In , tunkeat > > <tunkeat@g...> wrote: > > > > > > hi all, > > > > > > i am puzzled about the use of long-time predictor > > in celp. > > > i see in some implementations, there is a > > long-time predictor (ltp) > > > but in others, eg fs-1016, there isn't. if I > > understand correctly, > > > the ltp is used to compute the gain. but why is > > there a need > > > to do this computation when eventually, in the > > codebook search, > > > the gain and period of the pitch is "computed". > > thanks in advance. > > > > > > rgds, > > > tk > > > > > > > > > > > > > > > > __________________________________ > > To |