[Elecraft] Query for new DSP features

[David Woolley]

Do you have a reference for an algorithm that will do this?

Noise reduction is difficult because any effective noise reduction has 
to recognize what is signal, and ideally what is the part of the signal 
that matters to the human.  The hearing aid industry has been trying to 
do this for years, with limited success.

I think the sort of noise reduction we are talking about here 
essentially tries to decide which frequencies matter and which don't and 
eliminate the latter.  However, it has to do this when what it is trying 
to identify as signal is hidden by noise.

In practice, I think what these systems achieve is increased user 
comfort, rather than recovering signal from noise, as humans are 
probably still a lot better at extracting signal from noise than 
algorithms, but they get tired in doing so.

One consequence of selective filtering will be a reduction in total 
audio power.  I'd expect the total loudness to go down.  I guess you 
could then renormalise, and increase the signal power to bring the total 
power up to the same level.  However, most of are old enough to have a 
lot of high frequency hearing loss, so one may find that correction 
needed depends on specific hearing loss of the user and the original 
spectrum of the noise; one needs to renormalise the power as waited by 
the hearing sensitivity curve of the user.  I imagine you would need, at 
least, a parameter to determine the degree of renormalisation.

Also, the more aggressive you make this sort of noise suppression, the 
more likely it is to have false positives, and suppress important 
frequencies.  Also, the more aggressive you make it, the more you will 
get distortion as the result of modifying filter parameters on the fly.

Ultimately, though, the sort of noise that these systems are trying to 
remove is random in nature, so you can never be completely sure what is 
signal and what is noise.

(Hearing aids have a particularly difficult problem in that they are 
dealing with cocktail party noise, where the noise is the summation of 
lots of things that would, individually, be signals.)

(The ultimate noise reduction system would be one that recognized the 
speech and regenerated it, complete with characteristics of the original 
speaker.  However, doing that really well can only be done by looking 
ahead several seconds, to be able to interpret meaning from what 
follows, as well as what precedes.)

As a caution, I believe the K3 has two different noise handling 
strategies:  the one I am talking about here, and one designed to deal 
with impulse noise, where you simply cut out a short section of signal 
around the noise pulse.

-- 
David Woolley
Owner K2 01623

On 12/01/2019 04:38, Mike Lichtman wrote:
> I would like to see an improved Noise Reduction that doesn’t lower the volume or distort as much.