High freq whine

Hi all,

I wonder if you can help I have a recording and have been trying to remove high frequency noise for a while can you help please?

Have you tried the [u]Noise Removal Effect[/u]?

I can’t listen to your file or try it myself right now, 'cause I’m at work. It looks pretty bad, and it’s broadband noise, so I don’t have a lot of hope that you can remove the noise without nasty side effects. :frowning: Noise reduction works best when you have a constant low-level background noise… i.e. It works best when you don’t really need it.

FYI - Pros still record in soundproof studios with high-quality equipment because even with pro software, noise can be difficult or impossible to remove. But, with the reasonably priced equipment and a fairly quiet environment you can get nearly professional results at home, depending on what you are recording. (Noise comes form the analog hardware or the environment, so you don’t need high-end recording software.)

A method that always works for broad-band highish noise:
(copy the code in the nyquist prompt, under effects)

(defun red (snd mean_over)
  (snd-avg snd mean_over 1 op-average))
;; Change the "Nine" to a (odd) integer value of your taste
;; Higher values will also reduce the brightness 
(multichan-expand 'red s 9)

Thanks all.

Advice taken and the filter works great. I learnt something new today. Disconnect the broadband before doing recordings. Thanks for the help again one and all. Any extra tips for cleaning up recordings would be much appreciated.

Audacity all the way.

How do you mean that, What do you want to disconnect?
Is the source perhaps an old tape or a LP?
It sounds to me a bit like audio before dolby noise reduction or before a RIAA equalization.

If the noise I am getting on recordings is coming from broadband use while recording then I would just disable my network connection while record. This would save me the the hassle no?

That’s hardly necessary. It is in principle correct that all interference sources should be eliminated during recording. Those are however things like power suppliers or a screen that are to near to the computer/mic or just a leaky cable.

To me, it sounds simply to bright and sparkling. There may be an enhanced feature be enabled withing Windows like “Crystalizer” or Equalization for recording and or play back.
Those settings can be found in the sound control panel of windows (reachable via right click on the volume control in the taskbar or by typing “sound” in the search field of the start menu).
Click on the source (e.g. microphone) and on properties, the advanced tab may have some of these features enabled. look also into the playback section.

That’s all only a guess in the wild, there are a lot of Wiki pages dedicated to the recording setup.
I’m no expert in this field.

By the way, the code above could easily be transferred to a plug-in, if you don’t want to always put the text in the Nyquist Prompt. It would then appear in the effect menu under any name you like.

The high wind sound seems to be above 8500 Hz. It is a kind of feed back on the microphone. It gets clearer when the track is played at half speed.
There are a lot of cures for this case. Since it is apparently only speech, you can easily throw those higher frequencies away.
The best way is probably to resample all to 16000 Hz by changing the project sample rate (where normally 44100 is written).
This has the advantage that the file will also nearly 3 times shorter.
You can also use resample from the track menu (16000 Hz) or lowpass filter from the effect menu (48 dB, 8500 Hz).All this measures preserve the crispiness of the recording as opposed to my former general brightness reduction.

As to the cause itself, there seems to be any surface that produces this resonance frequencies. Try moving the microphone or the lap top around or use any tissue to damp the resonance (on the desktop or inside the microphone holder for example).
Hope that helps.

Thanks Robert

I will try all the suggestions. Thanks for the valuable help.

That is quite an unusual filter Robert.
I’m not currently at home so I can’t hear the posted audio sample, but looking at the spectrum of the audio sample I’m wondering if that filter (with multiple deep notches) is the most appropriate solution as I don’t see corresponding spikes in the audio.

It is a general white-noise filter. The notches change with the number provided.
A value of 4 or 5 cuts the upper frequencies enough to eliminate the offending wind noise. A little more reduces also some harshness in the consonants.
As I’ve said above, cutting the frequencies above about 8500 Hz is enough in this case and doing it per resampling makes certainly sense regarding the file size.
There may be other peaks in the spectrum but none that do resonate as much as the “elevation” frequencies (between 8000 and 9000 Hz, which tell if sound comes from above or below).

I was just wondering why not a low pass or shelf filter rather than N/2 notches at sample-rate/N, 2sample-rate/N, 3sample-rate/N…

You’ve rightly perceived that this is nothing but a lowpass filter. There are actually no notches apart from the stop band (for example, 2 would produce a notch at the nyquist frequency). It lies in the nature of the filter that the frequency response is fairly rippled.
I’ve used it because you’ve written “broadband noise” in your first post.
If we assume a signal + stationary,uncorrelated white noise, then the optimal filter is a moving average.
A lowpass filter is like convolving the signal with a weighting sequence like ‘0.1 0.15 0.5 0.15 0.1’ = 1 in total.
However, it makes no sense to weigh the samples at all if the additional noise is uncorrelated or totally randomly distributed. Therefore, the convolution sequence is simply e.g. ‘0.2 0.2 0.2 0.2 0.2’.
This gives the highest noise reduction for the general white noise case. If the actual distribution is known, another filter can be used (a matched filter). That’s what the Noise Removal effect does.
However, in the uploaded sample file, we have another kind of noise namely resonance which is most prominent during speech. It is therefore difficult to catch a noise profile.
The best filters are either areson, notch or the normal equalizer.
You have the advantage that you can say with a glance at the spectrum, which frequencies are offending. I find it a little bit querky to have to export the spectrum first to a text file and then to sort the values. I do therefore more often than not a judgement per ear and guesswork.

By the way, I find it sometimes useful to listen to just the residual after notching, i.e.

(diff s (notch2 s 8500))

It would be nice if the Audacity notch filter could return this residual audio too. Even better would be to be able to preview this audio–a proposal for a future implementation.

There are definitely notches for values greater than 2.
For example, if you filter white noise with a sample rate of 44100 Hz using a value of 5 you will get notches at 8820 Hz and 17640 Hz,
which correspond to 44100/5 and 44100*2/5

(defun red (snd mean_over)
  (snd-avg snd mean_over 1 op-average))

(multichan-expand 'red s 5)

Actually I think that was DVDdoug.

That’s true, of course. Let’s say it’s rather a combination between LPF and multiple notch filter (comb-like). Five is obviously a good choice for the sample file in question.
I usually use three, my old weak ears don’t mind a notch at 14700 Hz.

Oh yes, sorry.

By the way, I wasn’t criticising, just commenting that it is an unusual filter.
Do you know if it has a name?

It’s known as moving average or box filter.
It would profit from integrated oversampling (to keep the notch at Nyquist).
It is widely used for stochastic data smoothing to eliminate random influences in samples.
Better noise reduction is of course achievable with e.g a wiener filter that is rather based on least squares error or most likelihood, the kind of things that are built into cell phones.
Those filters are very hard and slow to implement in Nyquist.