Amateur radio receivers often have several types of audio filters, and a common one is simply called "Noise Reduction." Compared to notch filters, impulse filters, and adjustable bandpass filters, where the operation is fairly obvious, I have little idea how "Noise Reduction" works.

Some googling implied maybe it was simple smoothing, i.e. some kind of low pass filter, and it does sound like the high end is attenuated. But I can already do that with the adjustable bandpass filter. So I wonder if it is a little more sophisticated than that, perhaps with some non-linear or adaptive characteristics designed for voice on radio.

Any thoughts, and is there a term for this kind of filter that I could use for searching?

  • $\begingroup$ This depends on the manufacturer, and my information is old (vacuum tube radio old). Impulse noise reduction filters are often just called "noise reduction" on the front panel; they tend to be nonlinear with hysteresis (ideally they detect an impulse in a wider-band segment of the IF, then blank it out). For information from someone who uses current equipment you may want to migrate this question to the amateur radio stackexchange. $\endgroup$
    – TimWescott
    Commented Dec 4, 2023 at 23:12
  • $\begingroup$ @TimWescott Thanks. My radio also has an impulse filter which I understand better because it has parameters you can set. I think Noise Reduction came along with DSPs and I thought maybe audio processing people might have a guess as to what is does, as it would be useful beyond ham radio. $\endgroup$
    – gschro
    Commented Dec 4, 2023 at 23:38
  • 2
    $\begingroup$ There's a lot of overlap between the ham radio world and signal processing -- I'd be surprised if there weren't engineers on that group building radios, signal processing equipment, or both. $\endgroup$
    – TimWescott
    Commented Dec 5, 2023 at 0:36

2 Answers 2


In the olden daze, it was because the bandwidth was tight. 2.1 kHz crystal-lattice filter that had very sharp transition bands at 350 Hz and at 2450 Hz. Everything in-between passed and everything out of band was killed. And this was single-sideband, so we were not getting signal nor noise from the other sideband.

Then the other thing was a squelching circuit (or a "gate", this is a thing with hysteresis that I incorrectly originally thought Tim is alluding to) to mute the audio when no one was talking. That way the noise in pauses between words did not come through competing with silence in the content. It was only when there was a word being said that the in-band noise was allowed to also get through. But that noise had to compete with non-zero content.

  • $\begingroup$ Actually, no, I wasn't talking about squelch. There's a specific kind of filter designed to reduce noise from electrostatic discharge (i.e., lightning) and in days of old, the 21MHz Russian "woodpecker" over the horizon radar. Both of these show up as non-Gaussian, impulse noise, and much improvement can be had by detecting an impulse and killing it. It works best if it's applied before that narrow IF filter -- otherwise it's hard to distinguish between broad-band impulsive noise and actual signal content. $\endgroup$
    – TimWescott
    Commented Dec 5, 2023 at 0:39
  • $\begingroup$ hmmm. I'll make an adjustment. $\endgroup$ Commented Dec 5, 2023 at 1:03
  • 1
    $\begingroup$ @TimWescott after thinking about it, I can sorta see how, if you can properly detect the spike, you can remove it and replace it with some kinda interpolation of the signals for the very few samples where the spike was manifest. Linear interpolation or, better yet, two-sided LPC interpolation, appears to be a good way to do it. $\endgroup$ Commented Dec 5, 2023 at 18:50
  • $\begingroup$ I'd not be surprised if modern low rate vocoders are pretty much noise removal engines, as the optimization goal "make speech as comprehensible as possible, ignore noise" is the same for denoising and speech coding. Corollary, I think a sensible approach would probably be going with LPCnet as a starting point and tuning it for the noisy input scenario. $\endgroup$ Commented Dec 6, 2023 at 11:18

One common type of HF receiver Noise Reduction is spectral subtraction.

Spectral subtraction can be implemented using a non-linear algorithm which does voice activity detection, continuously (or at small time quantums) characterizing which frequencies are part of the human voice portion of the signal, and which are not. Then the algorithm uses adaptive DSP filtering to spectrally subtracts those portions of the spectrum (via FFT overlap processing) which the algorithm estimates/guesses are not part of the human voice input. If those portions of the estimated/guessed non-voice spectrum are noise, then noise is reduced.

This (sometimes) works because human speech has different statistical properties, in both the time and frequency domain, from many types of noise (AGWN, etc.) But if the estimation/guess based on these statistics is incorrect, then the result can contain some left over musical tones, or produce muffled speech.

AI/ML neural nets have also been used to detect and characterize speech versus noise in both the spectral and time domains, and determine which spectral portions to subtract and when.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.