Meaning of MCLT bins..., Modulated Complex Lapped Transform Options

[robiwan]

The MCLT produces "FFT-like" output, however, for a single frequency (with exactly N periods (N is integer) within window of size M), the MCLT produces a peak at two (2!) bins. What does this mean? I'm trying to use the MCLT for dynamic filtering of a signal, but then I'd need to estimate its spectrum, and to do that I need to know what the bins "mean".

TIA
/r

[HotshotGG]

The MCLT produces "FFT-like" output, however, for a single frequency (with exactly N periods (N is integer) within window of size M), the MCLT produces a peak at two (2!) bins. What does this mean? I'm trying to use the MCLT for dynamic filtering of a signal, but then I'd need to estimate its spectrum, and to do that I need to know what the bins "mean". 

I believe a "bin" in the loosest sense in terms of transformation is a statistical grouping of spectral coeffcients or a large number of them that would not be possible with a histogram.

This post has been edited by HotshotGG: Oct 16 2005, 23:55

[SebastianG]

The MCLT produces "FFT-like" output, however, for a single frequency (with exactly N periods (N is integer) within window of size M), the MCLT produces a peak at two (2!) bins. What does this mean? I'm trying to use the MCLT for dynamic filtering of a signal, but then I'd need to estimate its spectrum, and to do that I need to know what the bins "mean".

"FFT-like output" refers to the fact that you get two real values for a frequency bin and can reconstruct amplitude and phase with it. The difference to the FFT is actually the following:
- the MCLT's bin's frequencies peak at N/2 periods, N is an odd integer.
- windowing is sort of mandatory in a MCLT. So even if you produce a test signal with N/2 periods in the M sample window (N an odd integer) you still get more than one non-zero amplitude -- There's a peak at the expected bin, though.

If you don't have a good reason for using this transform, just do an FFT instead and reuse your current window function. It'll be much easier to code with pretty much the same effect -- plus you seem to be more familar with an FFT's output
;-)

Sebi

This post has been edited by SebastianG: Oct 17 2005, 02:23

[robiwan]

"FFT-like output" refers to the fact that you get two real values for a frequency bin and can reconstruct amplitude and phase with it. The difference to the FFT is actually the following:
- the MCLT's bin's frequencies peak at N/2 periods, N is an odd integer.
- windowing is sort of mandatory in a MCLT. So even if you produce a test signal with N/2 periods in the M sample window (N an odd integer) you still get more than one non-zero amplitude -- There's a peak at the expected bin, though.

If you don't have a good reason for using this transform, just do an FFT instead and reuse your current window function. It'll be much easier to code with pretty much the same effect -- plus you seem to be more familar with an FFT's output
;-)

Thanks Sebi, that explains a lot. The reason I want to use MCLT is precisely because it has no blocking artefacts, and I want to use it in time compression/expansion. However, to do that I need to change the phases so that when doing inverse MCLT, there'll be no artefacts. I'm not entirely convinced that this can be done though...

-r

[SebastianG]

Thanks Sebi, that explains a lot. The reason I want to use MCLT is precisely because it has no blocking artefacts, and I want to use it in time compression/expansion. However, to do that I need to change the phases so that when doing inverse MCLT, there'll be no artefacts. I'm not entirely convinced that this can be done though...

-r

I'm 100% sure that in this case there's no advantage over the FFT. There isn't 'cause you only use the outcome of the MCLT temporarly and transform it back for storage. The FFT is as good (possibly as bad!) for your puspose.

The MCLT and the (complex valued) FFT (using a similar window) don't differ greatly. The only thing is the already mentioned shifted bin frequency peak.

The cool thing about the MCLT is though that you can kill two birds with one stone. Example: Lossy audio transform coder. Those usually utilize the MDCT. By applying the MCLT you can do 1) FFT-like spectral analysis and 2) get the transformed MDCT samples (to be coded) for free (real part of the MCLT).

BTW: I believe you're better off doing cutting/blending in the time domain anyway.

Sebi