RLS and LPC comparison

Hi All,

I have implemented a 10th order Recursive Least Squares filter on
Matlab for speech and am comparing the weights thus obtained with the
weights using the LPC function. 

Then I am comparing the speech quality using PESQ MOS values for
various speech files.

I run my RLS for lambda=1 (forgetting factor)
and delta=.02 (regularization factor), which gives the best performance. 
Also, I am not using any windowing in the code

my questions are - 

1) I get PESQ MOS values like 4.47 and 4.45 for the estimated signal
using LPC. Are these values too high ?? Is there a limitation of PESQ
MOS in this case that I should be aware of?

2) For a particular speech file, the RLS weights and the LPC weights
match upto 3 decimal places, but the PESQ MOS value for LPC is 4.47
and that for RLS is 4.07. How come there is such a huge difference
between the two if the weights match almost perfectly? 

It would be great if someone can help me with this. 

thanks,
kartik

On Aug 6, 6:13 am, "kartikaye" <kartikay...@yahoo.com> wrote:
> Hi All,
>
> I have implemented a 10th order Recursive Least Squares filter on
> Matlab for speech and am comparing the weights thus obtained with the
> weights using the LPC function.
>
> Then I am comparing the speech quality using PESQ MOS values for
> various speech files.
>
> I run my RLS for lambda=1 (forgetting factor)
> and delta=.02 (regularization factor), which gives the best performance.
> Also, I am not using any windowing in the code
>
> my questions are -
>
> 1) I get PESQ MOS values like 4.47 and 4.45 for the estimated signal
> using LPC. Are these values too high ?? Is there a limitation of PESQ
> MOS in this case that I should be aware of?
>
> 2) For a particular speech file, the RLS weights and the LPC weights
> match upto 3 decimal places, but the PESQ MOS value for LPC is 4.47
> and that for RLS is 4.07. How come there is such a huge difference
> between the two if the weights match almost perfectly?
>
> It would be great if someone can help me with this.
>
> thanks,
> kartik

I dont know much about RLS but i have some questions.
What is the window size that you are using for LPC.

Difference between 4.07 and 4.47 is not huge.

On Aug 6, 8:52 am, rajesh <getrajes...@gmail.com> wrote:
> On Aug 6, 6:13 am, "kartikaye" <kartikay...@yahoo.com> wrote:
>
>
>
> > Hi All,
>
> > I have implemented a 10th order Recursive Least Squares filter on
> > Matlab for speech and am comparing the weights thus obtained with the
> > weights using the LPC function.
>
> > Then I am comparing the speech quality using PESQ MOS values for
> > various speech files.
>
> > I run my RLS for lambda=1 (forgetting factor)
> > and delta=.02 (regularization factor), which gives the best performance.
> > Also, I am not using any windowing in the code
>
> > my questions are -
>
> > 1) I get PESQ MOS values like 4.47 and 4.45 for the estimated signal
> > using LPC. Are these values too high ?? Is there a limitation of PESQ
> > MOS in this case that I should be aware of?
>
> > 2) For a particular speech file, the RLS weights and the LPC weights
> > match upto 3 decimal places, but the PESQ MOS value for LPC is 4.47
> > and that for RLS is 4.07. How come there is such a huge difference
> > between the two if the weights match almost perfectly?
>
> > It would be great if someone can help me with this.
>
> > thanks,
> > kartik
>
> I dont know much about RLS but i have some questions.
> What is the window size that you are using for LPC.
>
> Difference between 4.07 and 4.47 is not huge.

Also what is the order of the LPC and the sampling rate of the speech.

>On Aug 6, 8:52 am, rajesh <getrajes...@gmail.com> wrote:
>> On Aug 6, 6:13 am, "kartikaye" <kartikay...@yahoo.com> wrote:
>>
>>
>>
>> > Hi All,
>>
>> > I have implemented a 10th order Recursive Least Squares filter on
>> > Matlab for speech and am comparing the weights thus obtained with
the
>> > weights using the LPC function.
>>
>> > Then I am comparing the speech quality using PESQ MOS values for
>> > various speech files.
>>
>> > I run my RLS for lambda=1 (forgetting factor)
>> > and delta=.02 (regularization factor), which gives the best
performance.
>> > Also, I am not using any windowing in the code
>>
>> > my questions are -
>>
>> > 1) I get PESQ MOS values like 4.47 and 4.45 for the estimated signal
>> > using LPC. Are these values too high ?? Is there a limitation of
PESQ
>> > MOS in this case that I should be aware of?
>>
>> > 2) For a particular speech file, the RLS weights and the LPC weights
>> > match upto 3 decimal places, but the PESQ MOS value for LPC is 4.47
>> > and that for RLS is 4.07. How come there is such a huge difference
>> > between the two if the weights match almost perfectly?
>>
>> > It would be great if someone can help me with this.
>>
>> > thanks,
>> > kartik
>>
>> I dont know much about RLS but i have some questions.
>> What is the window size that you are using for LPC.
>>
>> Difference between 4.07 and 4.47 is not huge.
>
>Also what is the order of the LPC and the sampling rate of the speech.
>


Hi,

I am not using any windowing for LPC, am just giving the entire speech
file as input to compute the lp coefficients. LPC is of order 10 too, same
as RLS. I was asked to do this comparison so as to be sure that the RLS
code that I had written had no bugs in it, and that the results were almost
the same.

thanks,
kartik

Hi, 

I agree that the difference is not huge, but the fact that the weights
match up to 3 decimal places is what makes this difference appear big. I am
not sure why this is happening.

thanks,
kartik

On Aug 6, 8:59 am, "kartikaye" <kartikay...@yahoo.com> wrote:
> >On Aug 6, 8:52 am, rajesh <getrajes...@gmail.com> wrote:
> >> On Aug 6, 6:13 am, "kartikaye" <kartikay...@yahoo.com> wrote:
>
> >> > Hi All,
>
> >> > I have implemented a 10th order Recursive Least Squares filter on
> >> > Matlab for speech and am comparing the weights thus obtained with
> the
> >> > weights using the LPC function.
>
> >> > Then I am comparing the speech quality using PESQ MOS values for
> >> > various speech files.
>
> >> > I run my RLS for lambda=1 (forgetting factor)
> >> > and delta=.02 (regularization factor), which gives the best
> performance.
> >> > Also, I am not using any windowing in the code
>
> >> > my questions are -
>
> >> > 1) I get PESQ MOS values like 4.47 and 4.45 for the estimated signal
> >> > using LPC. Are these values too high ?? Is there a limitation of
> PESQ
> >> > MOS in this case that I should be aware of?
>
> >> > 2) For a particular speech file, the RLS weights and the LPC weights
> >> > match upto 3 decimal places, but the PESQ MOS value for LPC is 4.47
> >> > and that for RLS is 4.07. How come there is such a huge difference
> >> > between the two if the weights match almost perfectly?
>
> >> > It would be great if someone can help me with this.
>
> >> > thanks,
> >> > kartik
>
> >> I dont know much about RLS but i have some questions.
> >> What is the window size that you are using for LPC.
>
> >> Difference between 4.07 and 4.47 is not huge.
>
> >Also what is the order of the LPC and the sampling rate of the speech.
>
> Hi,
>
> I am not using any windowing for LPC, am just giving the entire speech
> file as input to compute the lp coefficients. LPC is of order 10 too, same
> as RLS. I was asked to do this comparison so as to be sure that the RLS
> code that I had written had no bugs in it, and that the results were almost
> the same.
>
> thanks,
> kartik

Then in my opinion you should compare segemental snr and not PESQ MOS..

Hey Rajesh,

thanks for the tip ... but I am not too clear about the segmental snr
thing, as I am not segmenting my speech file anywhere during rls and lpc.

kartik