Forums

Frequency warping

Started by VelociChicken November 12, 2008
Hello, I've been reading about frequency warping, but I'm not sure what's 
the best route to take.
It seems that using all-pass filters can't be reconstructed, and WFIR's are 
expensive, and there's warped wavelet techniques but I haven't bought the 
papers on this. Googling around gave me some mixed reports that CELP uses 
WPLC, which turns out to be an all-pass system that others said wouldn't 
work. So I'm a little confused...
What am I trying to do?
I'm experimenting with frequency scaling and have created a phase corrected 
pitch change using a vocoder like FFT bin multiply. All is fine, but I'm 
getting the problem of not enough bins in the lower part of the FT, and 
making the FT longer results in the usual smearing of the sound.
Is it worth my time doing a frequency warp to achieve more detail for lower 
bins, or is it not going to get decent results, or even work at all? I 
looked at phase locking, but it seemed people had problems with different 
kinds of audio sources, so this warping idea seemed a lot more elegant.

Thanks,
Dave



On Nov 12, 2:45&#2013266080;pm, "VelociChicken" <b...@yahoob.com> wrote:
> Hello, I've been reading about frequency warping, but I'm not sure what's > the best route to take. > It seems that using all-pass filters can't be reconstructed, and WFIR's are > expensive, and there's warped wavelet techniques but I haven't bought the > papers on this. Googling around gave me some mixed reports that CELP uses > WPLC, which turns out to be an all-pass system that others said wouldn't > work. So I'm a little confused... > What am I trying to do? > I'm experimenting with frequency scaling and have created a phase corrected > pitch change using a vocoder like FFT bin multiply. All is fine, but I'm > getting the problem of not enough bins in the lower part of the FT, and > making the FT longer results in the usual smearing of the sound. > Is it worth my time doing a frequency warp to achieve more detail for lower > bins, or is it not going to get decent results, or even work at all? I > looked at phase locking, but it seemed people had problems with different > kinds of audio sources, so this warping idea seemed a lot more elegant. > > Thanks, > Dave
Hi, Do you work in Matlab? There's a free toolbox for frequency warping here: http://www.acoustics.hut.fi/software/warp/ which might be a good starting point.
On Nov 12, 4:45 am, "VelociChicken" <b...@yahoob.com> wrote:
> Hello, I've been reading about frequency warping, but I'm not sure what's > the best route to take. > It seems that using all-pass filters can't be reconstructed, and WFIR's are > expensive, and there's warped wavelet techniques but I haven't bought the > papers on this. Googling around gave me some mixed reports that CELP uses > WPLC, which turns out to be an all-pass system that others said wouldn't > work. So I'm a little confused... > What am I trying to do? > I'm experimenting with frequency scaling and have created a phase corrected > pitch change using a vocoder like FFT bin multiply. All is fine, but I'm > getting the problem of not enough bins in the lower part of the FT, and > making the FT longer results in the usual smearing of the sound. > Is it worth my time doing a frequency warp to achieve more detail for lower > bins, or is it not going to get decent results, or even work at all? I > looked at phase locking, but it seemed people had problems with different > kinds of audio sources, so this warping idea seemed a lot more elegant. > > Thanks, > Dave
Dave A common brute force approach to 'get more bins' is to zero extend the FFT but keep the data blocking the same. If what you really need is more resolution you could use a larger FFT with larger data input block to reconstruct the lower frequencies and the current FFT for the higher frequencies. You need to be careful to keep the data blocks for the different sized FFTs on the same time centers to maintain phase correctly across them. Contemplating the complexity of this (although not all that great) has driven some implementers to consider wavelet transforms. Dale B. Dalrymple
> It seems that using all-pass filters can't be reconstructed
What do you mean by that ?
>On Nov 12, 2:45 pm, "VelociChicken" <b...@yahoob.com> wrote: >> Hello, I've been reading about frequency warping, but I'm not sure what's >> the best route to take. >> It seems that using all-pass filters can't be reconstructed, and WFIR's >> are >> expensive, and there's warped wavelet techniques but I haven't bought the >> papers on this. Googling around gave me some mixed reports that CELP uses >> WPLC, which turns out to be an all-pass system that others said wouldn't >> work. So I'm a little confused... >> What am I trying to do? >> I'm experimenting with frequency scaling and have created a phase >> corrected >> pitch change using a vocoder like FFT bin multiply. All is fine, but I'm >> getting the problem of not enough bins in the lower part of the FT, and >> making the FT longer results in the usual smearing of the sound. >> Is it worth my time doing a frequency warp to achieve more detail for >> lower >> bins, or is it not going to get decent results, or even work at all? I >> looked at phase locking, but it seemed people had problems with different >> kinds of audio sources, so this warping idea seemed a lot more elegant. >> >> Thanks, >> Dave
>Hi, > >Do you work in Matlab? There's a free toolbox for frequency warping >here: http://www.acoustics.hut.fi/software/warp/ which might be a good >starting point.
Hello, I don't know Matlab, it appears to call a few 'black box' functions, but thanks for the link as there is some good pointers there.
"dbd" <dbd@ieee.org> wrote in message 
news:9720b607-3e49-4976-b6cb-080f4b0797dd@k1g2000prb.googlegroups.com...
> On Nov 12, 4:45 am, "VelociChicken" <b...@yahoob.com> wrote: >> Hello, I've been reading about frequency warping, but I'm not sure what's >> the best route to take. >> It seems that using all-pass filters can't be reconstructed, and WFIR's >> are >> expensive, and there's warped wavelet techniques but I haven't bought the >> papers on this. Googling around gave me some mixed reports that CELP uses >> WPLC, which turns out to be an all-pass system that others said wouldn't >> work. So I'm a little confused... >> What am I trying to do? >> I'm experimenting with frequency scaling and have created a phase >> corrected >> pitch change using a vocoder like FFT bin multiply. All is fine, but I'm >> getting the problem of not enough bins in the lower part of the FT, and >> making the FT longer results in the usual smearing of the sound. >> Is it worth my time doing a frequency warp to achieve more detail for >> lower >> bins, or is it not going to get decent results, or even work at all? I >> looked at phase locking, but it seemed people had problems with different >> kinds of audio sources, so this warping idea seemed a lot more elegant. >> >> Thanks, >> Dave > > Dave > > A common brute force approach to 'get more bins' is to zero extend the > FFT but keep the data blocking the same. If what you really need is > more resolution you could use a larger FFT with larger data input > block to reconstruct the lower frequencies and the current FFT for the > higher frequencies. You need to be careful to keep the data blocks for > the different sized FFTs on the same time centers to maintain phase > correctly across them. Contemplating the complexity of this (although > not all that great) has driven some implementers to consider wavelet > transforms. > > Dale B. Dalrymple
Thanks Dale. I've tried zero padding to create very large FFT's but the quality is surprisingly not much of an improvement, probably for reasons you mentioned. I'm looking into Wavelets now - might take a while... : )