Hi, I'm trying to remove broadband noise from a recording in WAV format while retaining as much signal quality as possible, but I have a few questions about the artifacts it produces.

First, when I try to remove a large amount of noise from a particularly noisy part of the recording (this recording is a movie by the way), I end up with this weird oscillating mechanical sound that varies in frequency. If somebody could explain exactly what causes these mechanical sounds and how to fix or prevent them, I'd appreciate it.

Second, on parts of the recording with medium amounts of noise, sometimes just the voices sound a bit mechanical after the reduction. Is this the same effect as above or is something else causing this problem.

Third, on the parts where I can virtually eliminate the noise without serious consequences, I still have this issue where the voices seem to have some added echo or reverb. They don't sound as crisp, slightly muddy in fact, and they sound lower too. I've attempted to boost the higher frequencies to compensate a bit and I've made some adjustments to the sustain length, but neither of these methods seem to really solve the problem. I'd like if someone could explain this phenomenon to me and again explain how to fix or avoid the problem.

I've read some stuff about frequency-domain and time-domain aliasing. Is what I'm hearing the result of these effects? I realize that I should expect some signal degradation, but I want to know if there's more I can do to preserve the voice quality. Many parts have music too, and they may be influencing the noise reduction, but fortunately I can afford to cut out as much music as necessary because I wrote it and I can overlay the original tracks as I need to to compensate for the loss of music signal quality. I appreciate any insight on these matters.

- Blue Streak

I cant say I have a solution but I can tell you that I've also encountered almost exactly the artifacts you describe while trying to remove extremely high levels of "tape hiss" from old audio recordings.

the best way to go about noise reduction is to make many passes with very small incremental changes.

All broadband noise reduction I've ever come across uses a standard technique called "spectral subtraction", and the weird mechanical sounds you hear are a standard artifact of this method, generally called "musical tones". Musical tones are caused by the fact that the noise reduction is done by dividing the frequency spectrum up into lots (typically thousands) of narrow bands (using FFT) and doing the subtractions on each frequency band independently of all others. You inevitably end up with the odd frequency bands where the resulting amplitude after subtraction is much higher than the surrounding bands, and this leads to the musical tones. Because the particular frequency bands with the excess amplitude will vary from moment to moment, you hear the pitch of the tones moving around at random.

The way to remove the musical tones is to low-pass filter them out, but this is only feasible when dealing with low-quality speech: for music programme material, the tones are at far too low a frequency to be able to remove them without destroying the music. As dreamliner77 says, the best way minimise them is to do several light passes rather than one heavy pass.


I can't say for certain, but I suspect what you're hearing is indeed a milder form of the same thing.


If you're removing mainly high-frequency noise (eg. hiss), then you will inevitably lose some top end from the music. The reverb artifact is probably due to phse anomalies introduced by the subtraction and consequent FFT reconstruction of the waveform.


The only thing I know about "aliasing" concerns what happens during the sampling in A/D or D/A conversion. That has nothing to do with the artifacts you're hearing.

Wouldn't doing several light passes add up to a heavy pass? And if the reverb effect is caused by a phase distortion is it an inevitable consequence or is it repairable?

I can remove as much of the music as necessary because I will just overlay the track if I need to. What I'm really trying to preserve is the voice part of the signal.

- Blue Streak

Thanks for bringing this up, Blue Streak.
I'm dealing with the very same issue right now, but it's with recorded lectures rather than music.
So this caught my eye:
cliveb, what frequency do you recommend for a low-pass filtering of low-quality speech recordings to help remove some of these 'musical tones'? The files have been digitized as 22050 16-bit mono.

Thanks in advance!

~esa

Something I've noticed is that after removing noise, even in small increments, there are lines through the signal on a spectral view of the recording, as if somebody smeared the signal. Is there a way to repair the frequencies of the voice that were removed by the reduction process? They occur at all frequencies, not just high ones.

- Blue Streak

I disagree ... In my opinion, the best way is to use a De-Noiser that can produce residual output as a preview (like Sonic Foundry or Waves). I find this very handy since I can actually listen to what will be removed with my setting of choice and thus will be able to detect whether audible music content is affected or not.

JeanLuc,

I agree with using a tool that allows for residual output, but I stand firm in using multiple passes.

I do agree that 'less is often more' ...

No, because in between each light pass, you should take a fresh noise print. The musical tones are caused by specific isolated frequency bands in the noise print being quite large. After each light pass, the amplitude of the noise in each band will reduce.


Sorry, I don't know. I was only guessing that it may caused by phase errors. And if it is, my gut feeling is that these artifacts are not repairable.

It's impossible to say, because the frequencies and amplitudes of the musical tones will be determined by the nature of the noise itself and its interaction with the signal. All you can do is experiment, preferably using an editor with realtime preview capability on its filtering so you can adjust the filter while you listen. You'll never get a perfect result, so what you're aiming for is the best compromise.

I'll give it a shot - thanks!

Hey, I figured out my echo/reverb issue. I believe that I was using FFT with too many points. I was using up to 24000 points, and that caused the time-domain aliasing. Hence the pre and post echo effect. However, low FFT values don't discriminate between frequencies very well. Is the best solution right in the middle, like 4096 or 8192 points? Six thousand actually seems like a good choice; that's four Hz per band (it's a 48kHz recording).

On a related note, does anyone know of a way to reconstruct the frequency gaps in a signal created by spectral subtraction of noise?

- Blue Streak