Analog Subtract Audio From White Noise

The simple answer is "you can't." Audio+noise represent a thermodynamically more entropic state than just audio.

It's gone.

The white noise spectrum is all frequencies so you can't really filter it in analog.

Moving to software:

There exist single-ended noise reducers but they may harm more than they help and I use an old one that is out of production for the removal of narrowband noise. SFAIK, Adobe Audition offers the same.

You can look into CEDAR but it is quite expensive.

You are basically stuck with tweaking a low pass filter and prayer.

You can do something if you have for example 10 independent wax cylinders of the same archaic performance but you have to do it all digitally and adjust for the horrible wow and flutter of old kit.

In analogue you are stuck with some variant of tone controls = the only thing you can do well in analogue is a notch reject mains hum.

On a sunny day (Fri, 18 Dec 2015 17:38:19 +1100) it happened Mike Towner wrote in :

It all depends, 'audio', the noise level, but some info:

In general low level 'white noise' is not annoying.

The white noise is especially noticeable in the higher frequency parts, where it has a lot of energy.

IF you know (use an audio spectrum analyzer) where the maximum frequencies in your audio end, then you can low-pass directly above that to get rid of most of that perceived 'hiss'.

There are other ways too, using an expander - compander and playing with the thresholds. For example in Linux the 'sox' program can do that.

IIRC there is specialized software available to remove noise, no need for hardware at all.

Turn trebble down.

Perhaps I need to phrase the question in another way. It is really not about sound processing specifically.

Let's say I have a repetive audio signal consisting of three sinewaves mixed together.

Then I have white noise, which inherently also contains those same three frequencies.

Is there any way I can make the disappearance of the three frequencies within the white noise track the changing amplitude of the composite audio signal?

Extrapolating this to something like music, it would be making the frequencies within the audio track disappear from the white noise spectrum in proportion to their amplitude at any given moment of play.

Mike Towner

On a sunny day (Fri, 18 Dec 2015 21:28:06 +1100) it happened Mike Towner wrote in :

Do you _know_ these 3 frequencies? if so, are those amplitiude or frequency modulated and how fast? This determines the sidebands, the 'range' of frequencies you want to keep (or reject). If all this is known make 1 or 2 or 3 bandpass filters.

Are those frequencies related to each other?

How much is the amplitude of the noise relative to the 'tones' you have?

If you know the repetion frequency then you can use that to detect too.

I think at this point you better tell us what you are doing even if it is top secret and destroys the earth. Else it is just a guessing game.

What does the spectrum of a guitar or piano look like ? I would expect at any point of time there would be up to 6 or up to 10 tones active and their harmonics actively simultaneously. Taking FFT of the signal, setting those bins to zero that is not currently being played and then taking inverse-FFT to restore the signal might help.

Of course the attack and decay of a string might spread the tones so much, making such approach impractical, but at least these days we have the computing power to attempt such tricks.

On a sunny day (Fri, 18 Dec 2015 12:58:38 +0200) it happened snipped-for-privacy@downunder.com wrote in :

Oh yes, I was 'watching' some music (singer especially) with ffplay (Linux) and got very interested in the frequency modulation of the voice, amazing, I could not do it. ffplay on audio files does sort display a frequency versus time + amplitude.

Yes I have done the forward followed by reverse FFT, but for video only. In not-realtime you can do a lot of things, in realtime, the only case for doing things in hardware perhaps, it is much more difficult I'd think.

Your rephrased question still seems bizarre and I have no idea what you want.

However, if you want a noise reduction filter, Audacity has one built-in. It relies on you sampling a section of your input that has no audio, which it uses to characterise the noise spectrum. It then uses that to build a filter that reduces apparent noise in the audio signal.

No, I don't understand how that an work either, but it does, and quite well.

Clifford Heath.

Bayesian statistics or Kalman filtering are the keywords that will get you something along those lines but your question is still ill posed.

You can notch out any number of fixed frequencies. This sort of trick was done to prevent room resonance PA howl round although these days they tend to do it digitally by just shifting the frequency up or down a bit so that the resonance doesn't build up and feedback.

multiband noise gate

NT

Why do you want to do this?

What will you use it for?

Chris

Most people seem to think that you want to do noise reduction in a sigal, or fall back to the "why would you want that" question they pose when they have no answer to the original question.

When doing it in DSP, what you would probably do is this: Do an FFT on both the audio signal and the white noise signal (or skip the FFT phase and start with the frequency-domain representation of white noise), then you check the power level in the bins of the audio FFT, subtract that from the values in the corresponding bins of the noise FFT, then inverse FFT that result.

Is that a correct representation of what you want to do?

That is because for most applications you are generally trying to extract the very last bit of signal against a background of thermal noise that cannot be reduced further.

Why not just build an analogue white noise generator instead?

FFT it and replace all the complex components with a random value chosen to preserve total amplitude will reliably eliminate all signal from the resulting IFT but I cannot see why you would want to do it.

Try watching real music on a realtime waterfall FFT display and you will begin to understand the complexity of real audio signals plotted in terms of the amplitude and frequency as a function of time.

Why don't you read what Mike wants to do, instead of trying to fit it into a "most applications want..." model?

What he has said he wants to do makes no sense at all.

I have already answered what passes for his question in another thread. Bayesian analysis or Kalman filtering will allow adaptive filtering but there is no chance of doing it with analogue circuits.

His question is not well thought through and fundamentally clueless.

On a sunny day (18 Dec 2015 13:01:13 GMT) it happened Rob wrote in :

I am not sure even Mike knows that.

According to my newsreader you are the same Rob that posts to alt.nl.radio.zendamateur.gelicentieerd Right?

Unbelievable! Who are you to determine that what he wants to do makes no sense? Maybe he is trying to achieve some novel sound effect?

That would be a method, yes.

I think that position is not warranted. It was clearly explained what the goal was, that was not an impossible goal, and when you think everything you don't see the purpose for is "fundamentally clueless" that is a problem with yourself, not with those who ask a question.

If you know the frequences, you apply a filter to the noise.

The amplitudes of the tones don't matter, just the frequencies. The filter nneds to know the tone frequenies, not their amplitude.

If you don't know the frequencies, you need an FFT or something to determine the frequecies.

In real time with anlaog circuits...hmmmm not easy.

Better chance to do this in DSP

Mark

There are much simpler ways to turn a signal into white noise.

The hard bit is extracting tiny signals *FROM* white noise.