Re: 3D Audio -improving HRTF filtering (sherin tech )

Subject: Re: 3D Audio -improving HRTF filtering
From:    sherin tech  <sherin.tech@xxxxxxxx>
Date:    Thu, 6 Apr 2006 13:56:21 +0530

hi fred & Jose,

So,

If I am using rectangular window, ie, just taking M samples with a hop
size of again M.

for filtering using 128 point ( L ) filter ... I have to take Nfft >= M+L-1

i took, Nfft as 1024 ... and after the convolution I get back 1024
time samples ..

which, I am doing a 50% overlap and add with the previous 1024 samples ..

I am getting the o/p without any noise ..

Can we avoid this overlap and add ?? (i don't think)

regards,
sherin






On 4/5/06, Jose-Javier Lopez <jjlopez@xxxxxxxx> wrote:
> Hi Fred,
>
> Off course, any kind, is that I wanted to say (or rectangular windowing,
> that is the same).
>
> Windowing is indicated for signal analysis and other DSP things, but not for
> fast convolution using FFT.
>
> JJ
>
>
> ----- Original Message -----
> From: "Fred Herzfeld" <herzfeld@xxxxxxxx>
> To: "Jose Javier Lopez" <jjlopez@xxxxxxxx>
> Sent: Wednesday, April 05, 2006 7:15 PM
> Subject: Re: 3D Audio -improving HRTF filtering
>
>
> > Hi Jose,
> >
> > I agree that a Kaiser window should not be used, in fact for any kind of
> > accurate results no window at all shoud be used. Why did you single out
> > Kaiser?
> >
> > Fred
> >
> > Jose Javier Lopez wrote:
> >> Do not window input samples with kaiser !!!!!!!!!!!!!!!!!!!
> >>
> >> Review overlap-add or overlap-save algorithm in "Oppenheim -
> >> Distrece-Time Signal Processing".
> >>
> >> or look for code examples on Internet.
> >>
> >> JJ
> >>
> >>
> >> ----- Original Message ----- From: "sherin tech" <sherin.tech@xxxxxxxx>
> >> To: <AUDITORY@xxxxxxxx>
> >> Sent: Wednesday, April 05, 2006 11:38 AM
> >> Subject: [AUDITORY] 3D Audio -improving HRTF filtering
> >>
> >>
> >> Hi all,
> >>
> >> I am working on 3D-Audio based on HRTF filtering.
> >>
> >> My starting point is KEMAR data from MIT media labs, done by bill
> >> gardner and Keith Martin.
> >>
> >> I have implemented a basic by referering to the demo SGI code provided
> >> in MIT site.
> >>
> >> My first implementation is just time domain convolution , which works
> >> wothout any noise.
> >>
> >>
> >> But for efficient implementation I am trying to move to frame based
> >> processing, as in the SGI code provided in the MIT website.
> >>
> >> the steps involved in my new implemantation are ...
> >>
> >> 1) take 1024 number of samples.
> >> 2) Window those 1024 samples. (for this I am now using kaiser window of
> >> beta=16)
> >> 3) Take 1024 point FFT the input
> >> 4) Zero pad the 128 point HRTF filter to make it 1024 point.
> >> 5) Take 1024 FFT of HRTF Left ans Right data.
> >> 6) Multiply the spectrum in frequency domain.
> >> 7) Take 1024 point IFFT of the resultant spectrum.
> >> 8) Do a 75 % overlap and add.
> >>
> >>
> >> But a minor amout of noise is present in the o/p ....
> >>
> >> As a general instruction ... how can I improve the SNR of the o/p ?????
> >>
> >> I have tried , diferent type of windows, and overlap of 50 % also ..
> >>
> >> Any lights in to this ... ????
> >>
> >> regards,
> >> sherin
> >>
> >>
> >>
> >
> >
> > --
> > Fred Herzfeld, MIT '54
> > 78 Glynn Marsh Drive #59
> > Brunswick, Ga.31525
> > USA
> >
>