LPC Coefficients

Started by FruitfulToon January 30, 2005
Dear Readers,

I am currently working on some "speaker recognition" software which
will be using LPC coefficients for feature extraction. I have written
an algorithm but have no means by which to test it because of lack of
availability of existing LPC code.

I would greatly appreciate it if someone could send me some input
data, and output data for LPC (maybe 20 pieces?). I am using the
covariance method. My email is sorepharynx@hotmail.com.

Many thanks!
>>>>> "FruitfulToon" == FruitfulToon <sorepharynx@hotmail.com> writes:
FruitfulToon> Dear Readers, FruitfulToon> I am currently working on some "speaker recognition" software which FruitfulToon> will be using LPC coefficients for feature extraction. I have written FruitfulToon> an algorithm but have no means by which to test it because of lack of FruitfulToon> availability of existing LPC code. Why not feed some Gaussian noise through an AR filter, skipping over initial part of the output and feeding that data to your LPC estimator. The result should be close to the coefficients of your filter. Ray
Hi
You may  download the MATLAB code for LPC10 from the link below
http://www.cysip.com/LPC10-T.html
i am not sure if that meets your requirements. Note that it is suppose
to be for the LPC10 standard. You can atleast get a reference output i
guess.

-Nithin
FruitfulToon wrote:
> Dear Readers, > > I am currently working on some "speaker recognition" software which > will be using LPC coefficients for feature extraction. I have written > an algorithm but have no means by which to test it because of lack of > availability of existing LPC code. > > I would greatly appreciate it if someone could send me some input > data, and output data for LPC (maybe 20 pieces?). I am using the > covariance method. My email is sorepharynx@hotmail.com. > > Many thanks!
hi! I had implemented a speaker recognition software which too used LPC
for feature extraction. So what kind of an algo are you using (pattern
comparision?). What is the accuracy that you have got with it?

Hi,

I am using an autocorrelatoin method using the Levinson recursion. The 
algorithm I am using is adapted to java from some non-java code I found in a 
book. Do you know anywhere that I might find another basic LPC algorithm to 
test my own.

Thanks,

Tom
<kunal.yadwadkar@gmail.com> wrote in message 
news:1107347508.963669.291450@l41g2000cwc.googlegroups.com...
> > hi! I had implemented a speaker recognition software which too used LPC > for feature extraction. So what kind of an algo are you using (pattern > comparision?). What is the accuracy that you have got with it? >
Dear Sir,
I am trying to make speech processing task and i am trying to extract
the coeff. of sound signal in a .wav file.

I have read the wave file into a matrix

[x fs]=wavread('s000_001.wav');

x is 5005 * 1 matrix

fs is the sampling frequency.

fs = 12000

I think that i has to down the sampling frquency to 8000 using resample
function.

then i don't know what to do to get the coeff.

i am thinking now that will write the following code

while  j<= N_frames  no of frames
   A(j,1:11)=lpc(x(1:180,j),10);
   j=j+1;
end;

to pass the sound signal as frames to the LPC function the previous
loop i found on this newsgroup but i wonder do i have to use windowing
before as i read in digital processing of speech signals of Rabiner or
what ?

and what is the method to find the number of frames and choosing the
order of the LPC

What i had understand was that the number of coeff. = the order that i
use is that write or not ?

And as for Mr. Raymond Toy, u wrote:

to feed some Gaussian noise through an AR filter, skipping over
> initial part of the output and feeding that data to your LPC > estimator. The result should be close to the coefficients of your
filter. i don't understand what is the AR filter, what does AR stands for i know that AR stands for automatic recognition. There is somthing also i didn't understand in the text book of Rabineer i found that he feed the LPC with voiced and unvoiced signals but he said then that he use impulse signals as an input does he mean that he pass the sampled sound signal. does the unvoiced signal is the Gaussian noise. I am asking u because i found that u have experince in that field. Thanks Very Much, Best Regards, Marian Mounir Raymond Toy wrote:
> >>>>> "FruitfulToon" == FruitfulToon <sorepharynx@hotmail.com>
writes:
> > FruitfulToon> Dear Readers, > FruitfulToon> I am currently working on some "speaker
recognition" software which
> FruitfulToon> will be using LPC coefficients for feature
extraction. I have written
> FruitfulToon> an algorithm but have no means by which to test it
because of lack of
> FruitfulToon> availability of existing LPC code. > > Why not feed some Gaussian noise through an AR filter, skipping over > initial part of the output and feeding that data to your LPC > estimator. The result should be close to the coefficients of your > filter. > > Ray
>>>>> "marian" == marian mounir@gmail com <marian.mounir@gmail.com> writes:
marian> I think that i has to down the sampling frquency to 8000 using resample marian> function. I don't see why you have to down sample. LPC doesn't really care. marian> i am thinking now that will write the following code marian> while j<= N_frames no of frames marian> A(j,1:11)=lpc(x(1:180,j),10); marian> j=j+1; marian> end; marian> to pass the sound signal as frames to the LPC function the previous marian> loop i found on this newsgroup but i wonder do i have to use windowing marian> before as i read in digital processing of speech signals of Rabiner or marian> what ? Typically some kind of windowing is down before computing the LPC coefficients. This tends to the the LPCs more consistent, but biased. marian> and what is the method to find the number of frames and choosing the marian> order of the LPC The frame length is the important value. This depends on your sampling rate and on how fast you expect your LPC coefficients might vary over time. The order also depends on what you think the real LPC order is. There are ways of estimating the order, but I don't know how well they work in practice. marian> What i had understand was that the number of coeff. = the order that i marian> use is that write or not ? Yes. marian> i don't understand what is the AR filter, what does AR stands for i marian> know that AR stands for automatic recognition. AR here means Auto Regressive. Basically if you estimate the LPC coefficients a[k], you get an AR model of the signal and the AR filter has the transform, H(z): 1/H(z) = sum a[k]*z^(-k) marian> There is somthing also i didn't understand in the text book of Rabineer marian> i found that he feed the LPC with voiced and unvoiced signals but he marian> said then that he use impulse signals as an input does he mean that he marian> pass the sampled sound signal. marian> does the unvoiced signal is the Gaussian noise. I think Rabiner explains this. The speech is modeled as two sources. One is an impulse train that passes through a filter to produce speech. This is voiced. The other is Gaussian noise through a filter; this is unvoiced. These are approximations to the speech signal. The impulse train gives the periodic nature that some voice samples have. Ray