[Elecraft] K3S Noise reduction Test

[wa9fvp]

As I mentioned in my document 
 
“Human speech contains sinusoidal elements that the LMS algorithm adapts to and creates a filter around the speech waveform.  The waveform constantly changes and to conform to the human speech patterns, the LMS filter must continuously change the filter’s shape.  Under these conditions, it is too difficult to characterize the K3/K3S NR adjustments.  
 
That’s why I used a single tone test.
 
Human speech contains sinusoidal along with non-sinusoidal elements.  TH, S and W are a few non-sinusoidal elements and are more analogous to white and pink noise.  By attacking these elements, an aggressive NR reduction scheme can rendered speech unintelligible.  Other languages have an abundance non-sinusoidal speech and the question is, how will a NR scheme treat it?  In the telecom engineering lab where I worked, to test eco-cancelers, we actually used recordings of other languages.  Using tools that I have at my disposal, my test was to simply characterize the K3/K3S NR system.  
 
Noise reduction is magic but the bottom line is, how much does it reduce noise and is it so aggressive that is makes human speech un-intelligible?  The most difficult part is to find the correct parameters that can satisfy both conditions.  
 
-- 

Jack WA9FVP- http://www.willcoele.com/radio_repair  <http://www.willcoele.com/radio_repair> 

The Radio Reclamation Center (815) 463-9365
 
From: Grant Youngman-2 [via Elecraft] [mailto:ml+s365791n7631926h22 at n2.nabble.com] 
Sent: Tuesday, June 20, 2017 2:46 PM
To: wa9fvp <repair at willcoele.com>
Subject: Re: K3S Noise reduction Test
 
You might want to look at actual (S+N)/N data.  I’m not sure the visual “test”  tells the full story, since these are complex waveforms. 

I did extensive measurement several years ago on the NR functions of the v1 and v2 firmware versions of the TenTec Orion (for some reason, noise reduction not being “magic” surfaces everywhere as a contentious issue), and the measured data did not always match what things ”looked like”.   Not saying they don’t here, but it’s improvement in (S+N)/N that’s the end objective of NR regardless of what the screen shot looks like. 

SpectrumLab can measure this directly.  http://www.qsl.net/dl4yhf/spectra1.html

> On Jun 20, 2017, at 3:13 PM, wa9fvp <[hidden email]> wrote: 
> 
> If you look at my data, there’s very little noise level difference in the delay settings. 

Grant NQ5T 
K3 #2091, KX3 #8342 







-----
Jack WA9FVP

Sent from my home-brew I5 Core PC
--
View this message in context: http://elecraft.365791.n2.nabble.com/K3S-Noise-reduction-Test-tp7631913p7631928.html
Sent from the Elecraft mailing list archive at Nabble.com.