[Elecraft] [K3] Noise Reduction Recording

[David Woolley]

Trevor Smithers wrote:
> 
> What you really need is a method of identifying the actual ssb speech waveform, extracting the noise 

More or less by definition, if you have mix of an unknown signal and 
noise, you cannot separate the two.  That's because noise is random and 
therefore unpredictable and the signal is unknown, so also unpredictable.

To remove some of the noise, you must make assumptions about the signal. 
  For speech, that probably means assuming that only discrete formant 
frequencies are present (but that isn't valid for unvoiced sounds and 
particularly for s and h sounds - the loss of these may be why people 
find noise reduced signals to be muddy), identify those frequencies, and 
then attenuate all the other frequencies.

If the signal is close to the noise, you will not be able to identify 
the important frequencies properly, and therefore not be able reduce the 
noise this way.

More drastically, you can try and construct a set of voice tract model 
parameters, and then synthesize the speech from those (this is what GSM 
does, but for bandwidth reduction rather than noise reduction, so it 
assumes a reasonable starting SNR).

Most drastically, you use speech recognition and have the noise reducer 
re-speak the words.  Currently the best tool for doing speech 
recognition in the presence of noise is the human brain, though, so the 
best way of getting noise reduction is to get someone else to listen to 
the signal and repeat it to you!

Generally, noise reduction has to be more of an operator comfort thing 
than really improving the signal.  If the noise is so strong that you 
cannot make out the speech, the noise reduction isn't going to be able 
to find it anyway.

-- 
David Woolley
Emails are not formal business letters, whatever businesses may want.
RFC1855 says there should be an address here, but, in a world of spam,
that is no longer good advice, as archive address hiding may not work.