Recently, I read about noise reduction with artificial intelligence. AI is quite fashionable and, in my view, is seen too much as a panacea.
Here are some thoughts based on some theory.
If there is “pure noise” (white noise) then you cannot filter out that noise. AI or no AI, it is impossible. If it were, then all theory books can go in the dustbin. White noise is random and therefore it is impossible to predict what the waveform will look like in the future. It is different every time.
When there is interference, for example from “man made” sources, it is possible to reduce the interference in certain cases. A simple example is a carrier wave. If it is stable, then it is simple to predict what the next period will look like. This is because it is a copy of the previous period. If you can reconstruct the carrier wave (digitally), you can add it in counterphase to the signal + interference and the signal remains (more precise: you subtract the interference). You can also reduce interference from a carrier with a notch filter, but in that case the desired signal is affected by the filter properties. Many transceivers with DSP have such an (auto-)notch filter and if a carrier is strong, human hearing can no longer separate it properly (our hearing is very good at “signal processing” but everything has limits). But for SSB, it is certainly useful.
3. Phase box
The counter-phase trick is used in X-phase and similar noise reduction devices. There is no AI in those, but if you can pick up the interference as cleanly as possible, for example by placing an auxiliary antenna close to the interference source, it can work. The phase box is tuned so that the interference signal is added with equal amplitude and in counterphase to the signal + interference from the “main antenna”. For a single source of interference, this works pretty well, provided the auxiliary antenna picks up the interference properly. By the way, it is also a method of diversity reception.
4. Recognising interference
Without a phase box and reception of the interference signal, it becomes very tricky. Because the AI software would have to constantly monitor and adjust the interfering carrier, because even the smallest variation is enough to cause AI to fail. Even if you have a transceiver coupled to a stable (10 MHz or so) standard, it is still questionable if it will work.
As the waveform of the interference becomes more complex, it becomes increasingly difficult to recognise it as interference and the software will make less and less of it. Even if the (AI) software is “trained”, the less the interference “repeats”, the harder it will become to remove it. See also footnote (a).
5. Making use of signal properties
The desired signals themselves may contain properties, which allow the signal to be enhanced. This is often referred to as noise reduction. A CW signal is an on/off switched carrier wave and because it is a carrier wave, the spectrum is narrow. Moreover, a point or dash lasts for a number of periods and these two properties can be used to apply spectral averaging. With FFT, the spectrum of successive “slices” of the signal in the time domain is calculated and the spectra of these slices are summed. Then the calculated sum is transformed back to the time domain with the reverse FFT operation. Noise is random and therefore averages out, but because the spectra of the (CW) carrier wave are stable, the signal “stays” while the noise subsides, so to speak. In my experience, weak CW signals are better with “Noise reduction” of my Icom 7600. I don’t set the reduction too aggressive, because in that case, artefacts become dominant and it does no longer improve. I set it to about a quarter.
The JT modes (as also FT8) also make grateful use of spectral averaging. The slow transmission makes this possible. CW is often “slow enough” for averaging. With fast CW, the “noise reduction” will not work as well. With SSB, by the way, I do not notice any difference in intelligibility. By the way, do not use the transceiver noise reduction for FT8 etc. as it will only deteriorate. You can see on the waterfall graph that the signal is “smeared out”. The software itself takes care of the noise reduction.
There are indeed possibilities to improve signals (somewhat). So in certain cases it works, even if the interference is stronger than the desired signal, like with a phase box (or DSP equivalent). But real magic?
One of the tricks of manufacturers of electronics to pass “CE approval” is to frequency modulate the switching frequency of switched circuits with noise (think switching inverters, switching power supplies etc). Modulation with noise “transforms” interference into noise and my take is that no AI will be able to improve things.