> Allow me to propose a test of the difference between
> autocorrelation and a narrow filter as the noise reduction
> process. Check a frequency with multiple weak signals, maybe
> a multiple tone HF data link. Note that if the NR works it
> enhances all the signals to noise while a narrow tracking
> filter would enhance only ONE of those signals.
Actually, not necessarily.
Autocorrelation (and also the cross correlation matrix) IS used in one form
or another in establishing the error term between the de-noising filter's
output and the desired output (the denoised signal).
Most typical adaptive noise reduction algorithms (LMS, leaky LMS, steepest
descent, LRS, pick one) implement an adaptive Wiener filter. It can be an
IIR filter, or an FIR filter, but most usually an FIR form filter of some
number of taps (driven by acceptable processing delays, signal/noise
characteristics, whatever).
The gain of a Wiener filter at any particular frequency is a function
(roughly) of the SNR at that frequency (this is a power spectrum ratio, not
the usual SNR). So if the SNR is high, the gain fo the filter is high
(relatively), and if the SNR is low, the gain is low.
As a result, the Frequency Domain characteristic of a Wiener filter
operating on a signal with multiple spectral peaks (RTTY, SSB, for example)
is a filter response that broadly follows the signal spectrum. For example,
the Wiener response to a two-tone RTTY signal (depending on adaptions rates,
and number of taps in the FIR filter, whether the noise is really Gaussian
and stationary, all that "etc" stuff) may look more or less like a 2-humped
thing. The response to a voice signal is going to be something that
generally has a "bandpass"-like characteristic around the major formants.
(By the way, this was quite visible in the v1.xx Orion output spectrum).
You can look at it as "notching" those spectral ranges that contain only
noise and no signal, or as a bandpass around major signal spectral ranges
... Or something. It doesn't really matter since the effect is the same.
I think we're all really on the same wavelength here, but using different
vocabulary or talking about time-domain vs. frequency-domain, and just
talking around each other. Of course none of this says anything about the
actual filter construction in the Orion (or Yaesu or Icom) radios since I
don't think any of us has had the privilege of seeing the filter formulation
or the code or the MATLAB simulation or whatever.
The reference for these comments, in the event anyone cares, is "Adaptive
Digital Signal Processing and Noise Reduction", Saeed V. Vaseghi, Third Ed.
Chapter's 6,7.
Grant/NQ5T
_______________________________________________
TenTec mailing list
TenTec@contesting.com
http://lists.contesting.com/mailman/listinfo/tentec
|