Sign in
Search speechcoding
Subscribe to speechcoding
speechcoding by Keywords
ACELP | ADPCM | AMBE | AMR | AMR-NB | CELP | Codebook | DTMF | G.723 | G.726 | G.729 | GSM | Interpolation | LPC | LSF | LSP | MELP | PCM | Perceptual | Pitch | PSOL | QCELP | Quantization | SMV | VAD | Vocoder
Ads
Discussion Groups
Discussion Groups | Speech Coding | [Fwd: Re: Re: Fast convolution]
Technical discussions related to Speech Coding (all itu and other vocoders, ACELP, CELP, AMR, etc)
[Fwd: Re: Re: Fast convolution] - Jeff Brower - Feb 24 22:59:00 2005
|
-------- Original Message -------- Subject: Re: Re: [speechcoding] Fast convolution Date: 25 Feb 2005 03:31:46 -0000 From: "venkat ramanan" <> Reply-To: "venkat ramanan" <> To: "Jeff Brower" < hai Thanks for reply. I have a difficulty in taking fft. I have to take a 40 point fft. (all the inputs are always real only). So i padded 24 zeros and make in 2 power and take 64 point fft for both impulse response and the input. and then multiplied the both and take inverse transform. But the result is not much accurate. (i am using fixed point arithmetic). Can anybody suggest a very good way for implementing a 40-point FFT? (I have some material for mixed radix fft and that not clear). Venkat Mistakes are not end of the world but repeating them is |
(You need to be a member of speechcoding -- send a blank email to speechcoding-subscribe@yahoogroups.com )
Re: Fast convolution] - Shree Jaisimha - Feb 25 0:46:00 2005
|
Venkat,
To add on to this from Booshan's mail and Jeff's comments, here are some points to
consider.
You mentioned you are doing a 40 point FFT. Please note that any type of fast convolution
has some disadvantages.
They are as follows:
(i) Increased Memory locations on your DSP RAM to do the block processing, example :
Y(w) = H(w)X(w)
(ii) There is also inherent processing delay. You need to store the Frame and make sure
its full before you do the frequency domain multiplication as compared to time domain
convolution which is sample by sample.
With regards to computational savings as Booshan pointed out, with fast convolution
its O(4*([3N/2 lg(base2)N + N)]) while normal time domain convolution will use up 2N - 1
assuming equal length sequences of length N.
Fast Convolution is ONLY faster if the following inequality is satisfied:
6Nlog(base2) N + 4N < N^2
In your case N=40, please do the computation above and see if the equality is satisfied.
If not fast convolution is not for your end application as its not going to help in terms of
computational savings. However since N is a power of 2, you need to append zeroes and the
nearest N=64.
Even for Fast Hartley there has to be an inequality to be satisfied before its
efficiencies that make it "Fast" are realized.
By the way are you from PSG Tech in Coimbatore? :) One of my good friends is from there.
Keep up the good work. Valka Valarka.
Hope this Helps.
Sincerely,
Shree Jaisimha
In God We Trust.
Jeff Brower <j...@signalogic.com> wrote:
Shree Jaisimha
|
(You need to be a member of speechcoding -- send a blank email to speechcoding-subscribe@yahoogroups.com )