Analog Subtract Audio From White Noise

Pick a freqeuncy band that is not close to the three sine waves. Feed that "inverted" to a VCA such that if there is energy in that band, the VCA drops the output level dramatically.

This is analogous to a compressor with a sidechain, fed by a bandpass filter connected to the sidechain.

--
Les Cargill

As I read the second formulation of the question, it's like a set of DTMF style tones ... "TDMA" with white noise - at time T, you either have tones or noise ( or presumably silence ).

YMMV.

--
Les Cargill

I don't think the noise is mixed with the signal. They are separate. So this is *not* a noise filtering issue.

He wants to filter the noise at the frequencies of the separate signal. Not sure why this would be useful, but the remainder would be noise with gaps in the spectrum corresponding to the frequencies of the original signal without the noise. He even wants the filter to set the degree of attenuation to correspond to the amplitude of the signal at that frequency.

--

Rick

I can describe this with analog filters. He talks about the signal track and the noise track, so I assume they are not combined, but rather separate.

The signal can be analyzed by a fine grained frequency analysis, perhaps a series of bandpass filters. The white noise is also filtered through an equalizer type function with the same frequency resolution. The signal strength in the bins from the signal source are used to control the gains settings of the white noise equalizer.

This would be a very large circuit, but all of the parts are not hard to design or build. In fact, the same analog circuit comprising both the signal bandpass filter and the white noise equalizer channel could be made as a module to be used one for each frequency bin.

--

Rick

Thanks. A few people, including yourself and Rob, have interpreted my question correctly.

My apologies for any lack of clarity. I want the composite frequencies of the audio signal (voice, music, sound effects, etc.) substracted from the white noise spectrum in real time.

This is based upon the premise that the noise contains all frequencies when sampled over close intervals. Thus it contains, in undifferentiated form, all frequencies of the audio signal.

The object is to make the frequencies contained within the audio signal disappear from the noise. This creates holes in the otherwise "continuous" noise spectrum which are modulated by the audio signal.

It is the opposite of adding the track to the white noise.

How can this be done with either analog circuitry or perhaps commercial music editing software?

Mike Towner

OK this finally makes some sense. You've got one big problem Mike, and that is that the noise has all the frequencies, but at some unknown phase. (And the phase changes with time... it's noise after all.) So you can't subtract some fraction of the sine wave and get a hole.

The best way to make a hole in a noise spectrum is to either notch filter a white noise source, or create your own pseudo-random noise source with DDS and leave out the frequencies you don't want.

George H.

Still, this whole thing is typical: Someone asks for a circuit doing something based on a preconceived and probably misguided idea. The crux of the matter is really: What are you trying to do? Is it some sound effect? What would you expect to hear then? Is it something else?

Can you just take a step back and explain what purpose such treatment is supposed to accomplish?

Jeroen Belleman

Huh? White noise is indistinguuishable from white noise, just use a pure noise source.

So you want noise only in the bands where audio isn't? you'd need to sense the frequencies present in the audio and then attenuate the matching bands in the noise.

this means a bucket of analogue circuits or some hairy DSP code,

--
  \_(?)_

Hi Mike

Are you trying to design a voice scrambler / descrambler system?

Reading your description (and the rest of this thread) I can't help thinking that your proposed filter device, when applied to itself, actually would seem to invert its own function - and while the signal "in the middle" would mostly sound like white noise to human ears and therefore be unintelligible, after the signal was passed through this "filter" again, it could be at least semi-recovered.

However, there's a catch, and the catch is that you will lose all the phase information from the original signal. Running your proposed filter function, twice in a row, on some audio signal would result in a signal with the same amplitude-frequency characteristics, but with a completely randomized phase of each and every frequency component. I think though, that such a signal, if it were speech, and if bandwidth limited from about 200 Hz to some kHz (where the frequency range of human speech is and also where the phase information is not essential for human hearing), would be understandable reasonably well.

If your intention is completely unrelated to this, then considerably more details on the required time and frequency resolutions are needed to see if what you are asking for is feasible.

Frankly, I have no idea how to do this in an analog circuit of a realistic size. It's simple in theory - rickman just gave the full description (a filter bank with RMS level outputs driving the gain controls of a matching equalizer) - but this circuit is so unwieldy that even with the best of intentions I wouldn't consider it practical.

If your frequency resolution needs are very low then it would be realistic, but then "very low" would probably mean countable on the fingers of one's hands. But then, SIGSALY (look it up on wiki) only provided frequency discrimination in 10 bands, amplitude discrimination in 5 levels, 25 updates (samples) per second, and no phase data at all. And it was intelligible enough for practical use for security needs.

However, if you do the filtering in software on a reasonably fast DSP, it should be possible to get many more frequency bands and a faster update rate than that. An FFT-based solution running at some 200 FFTs per second or so, could probably even (in a sort of OK enough way) approximate natural human speech.

Regards Dimitrij

Well, this is sorta like a multiband noise gate, run backwards in a way.

1) Do an FFT of the noise signal to determine the power and phase of the noise in each band. 2) Do a second FFT (same bucket sizes and frequencies) of the desired audio signal.

(3) Modify (reduce) the power present in each bucket of the noise signal's FFT, based on the amount of power present in the corresponding bucket of the audio signal's FFT. You could do anything from "force the noise bucket to zero whenever the audio bucket exceeds a specific threshold", to "attenuate the noise bucket by M dB, for each N dB of power in the corresponding audio-signal bucket", or whatever you like,

(4) Run an inverse FFT on the noise signal's spectrum to generate the filtered noise.

In order for the effect to be really audible you may want to zero or attenuate the noise power not just in the "matching" buckets, but also in the buckets surrounding them... maybe as much as a third-octave worth.

Hi Mike

Come to think of it, you'll probably want the whole thing to work in real time, smoothly, and in a way that preserves the timing information of each frequency component. Phase may not matter much, but at least the onset timing of each frequency "note" should be preserved and subtracted from the noise at the correct instant. Unfortunately the FFT approach won't really do that, especially for low frequencies, where the fact that the phase information is missing will be noticeable.

Please strike what I've just written about using FFTs. With FFT alone the overall result of your filter is likely to sound very "jumpy". To implement your proposed function digitally one would more likely need lots of FFTs with windowing or a wavelet transform instead. The wavelet transform would preserve the timing of each amplitude change at each frequency. So, if you need a filter that works smoothly without "jumps" (dictated by the beginning and end of each FFT block), you'd be better off using the fast wavelet transform for the audio signal analysis part, and the inverse fast wavelet transform for the synthesis part of the resulting noise "signal". Unfortunately, in this case, the processing required to "apply" the wavelet representation of the signal part to the corresponding wavelet representation of the noise part is going to be more complicated than with FFTs. It would also likely need much more processing in the wavelet domain too. In the frequency domain a simple magnitude calculation and attenuation would have been enough, in the wavelet domain one would have to take timing into account, which would make it decidedly non-trivial. Unfortunately that's also where my basic understanding of wavelet transformations ends, so you'd need to ask for help an expert if you want to go that route.

Regards Dimitrij

To understand what's going on you need a good 1000 frequency channels and rapidly falling bars. Consumer hifi displays are misleading in this respect.

NT

** Audio frequency shifting is better and more cheaply done with analogue c ircuitry. A pair of multistage phase shift networks, a couple of multiplier ICs and a LF sine wave oscillator are the ingredients. I designed a high quality one for PA system use using TL074s and AD633s. The design was published and a kit of components inc PCB sold for about A$50.

DSP versions may exist, but I have never successfully tracked a REAL one do wn.

Such frequency shifters are a controversial devices, those who have never a ctually used one are often highly sceptical or even hostile to the very ide a while those who have used one are normally very enthusiastic about the re sults.

There was quite a bun fight here on this newsgroup about 10 years back conc erning the audibility of small frequency shifts, like +5Hz, on music progra mme.

So I posted a couple of music files ( shifted v unshifted) that cleared the matter up are far as rational folk were concerned.

.... Phil

Thanks to the inupt so far I am beginning to understand the nature of the problem.

The description above seems like the most viable approach.

Can anyone recommend a software package that can do this?

Preferrably one that has these functions built-in rather than needing to be coded.

Many thanks,

Mike Towner

** That is a miscomprehension.

The notion that white noise is a mix of "all frequencies" is a maths fiction - white noise is simply a randomly varying voltage. IOW, the value seen at some future point in time is unpredictable.

The maths fiction is that it can also be seen as the sum of an infinite number of sine waves at an infinite number of frequencies - the implication being that each one has an infinitely small amplitude.

So it's removal would be undetectable.

.... Phil

Since the MP3 file format describes frequencies and associated amplitudes at evenly spaced time intervals through the file, you should be able to take an MP3 file and process the data to produce another MP3 file containing the spectral inverse of the file, which appears to be mostly what the OP wants to do.

The problem is with the question and the premise that doing this would be useful. It isn't and if you want to see why mock up a pseudorandom noise generator that omits a few spot frequencies in software.

Or feed a white noise generator into a notch reject filter.

But at a random and shifting phase.

No it isn't. If were that simple you could just subtract the track from the white noise. White noise is by its nature noisy.

The short answer is that it can't mainly because doing this would have no useful purpose. White noise with a few spot frequencies missing sounds well pretty much like any other white noise. You could implement a tracking series of notch reject filters but why would you bother?

Even lopping great chunks out or biassing it to pink noise with a low pass filter it will still just sound like noise.

What are you trying to achieve?

--
Regards, 
Martin Brown

It is true after fashion provided that you allow that it is a mixture of sin *AND* cos with uniform expectation of amplitude but random phase.

The white noise will contain some of the signal channel frequencies in band which in conventional detection determines signal to noise.

That will be roughly what the OP will find if he implements a notch reject filter (or bank thereof) and feeds white noise into it.

He could save a lot of time by not bothering.

--
Regards, 
Martin Brown

Aren't we a bundle of joy today?

--

Rick