Hiss/Noise Removal

[brispuss]

I have a similar issue as that for meeshu's post regarding removal of hiss/noise from the audio track of a video.

Although I don't own Izotope, I have access to it for audio editing.

Using Izotope I've tried to remove the hiss/noise by using Deconstruct, and later tried using Spectral De-Noise.

First, tried setting to high noise dB removal for one pass in rendering, then tried setting to low noise dB removal for multiple passes in rendering.

But the results are usually about the same. The noise is considerably removed, but there are artifacts caused after the rendering which distorts the wanted remaining audio (music and effects). The distortion sounds like bubbling or a watery type of sound, and although not too bad it is noticeable.

The sample has wanted background music and some effects as well as the noise.

How to remove this noise without causing any distortion of the wanted sound (music and effects)?

Audio_trimmed_to_noise.flac

[davexnet]

Did you find a second of just noise to take a noise print?

2 passes of Deconstruct

[timtape]

...How to remove this noise without causing any distortion of the wanted sound (music and effects)?

I suspect it's impossible. Listen to a range of online audio noise reduction demonstrations from different vendors and see if you can find an example with such a weak signal under such heavy noise (more noise than wanted signal), which has then been successfully denoised without audible artifacts. I've been looking for some years and have not found one. Certainly denoising tools have improved significantly since they were first introduced professionally in the late 80's but not to the point of dealing with examples like this.

The standard first step is to find a better example/ transfer of that particular recording.

[brispuss]

Did you find a second of just noise to take a noise print?

2 passes of Deconstruct

No. I forgot to mention that I couldn't find a section of audio with just the noise to use as a print or mask.

Your example appears to sound better, but there is still noticeable noise. Thanks anyway.

[brispuss]

...How to remove this noise without causing any distortion of the wanted sound (music and effects)?

I suspect it's impossible. Listen to a range of online audio noise reduction demonstrations from different vendors and see if you can find an example with such a weak signal under such heavy noise (more noise than wanted signal), which has then been successfully denoised without audible artifacts. I've been looking for some years and have not found one. Certainly denoising tools have improved significantly since they were first introduced professionally in the late 80's but not for dealing with examples like this.

The standard first step is to find a better example/ transfer of that particular recording.

Thanks.

The first thing I actually did was to find a section of the track with just noise only to use as a print or mask, but I couldn't find any noise only sections, unfortunately.

Finding alternative better sources may be possible, but may take quite some time(?)

[timtape]

Thanks.

The first thing I actually did was to find a section of the track with just noise only to use as a print or mask, but I couldn't find any noise only sections, unfortunately.

The "noise print" feature or requirement on some older denoiser tools seems a lot less common these days.

Finding alternative better sources may be possible, but may take quite some time(?)

It depends on the recording. If it was a publically sold record and you know the name of the song and the performer, an online search should yield something. Someone with a knowledge of that genre and period of recorded jazz music might recognize the tune and band.

It sounds like an old 78 record copy that was extremely worn out and/or inexpertly played. Possibly a good vertically cut disc such as an Edison Diamond Cut disc but unknowingly played as if it was the more common laterally cut disc.

https://en.wikipedia.org/wiki/Vertical_cut_recording

[smartel]

try a stem splitter

[pandy]

Blind noise removal is not easy task - perhaps there are some "AI" novel methods working better than classical approach.
FFmpeg offers some audio denoising methods (afftdn , anlmdn , arnndn).
For neural method check https://odan.github.io/2023/08/20/reducing-background-audio-noise.html

For classic approach(spectral deconvolution with help of FFT) i use Adobe Audition (former CoolEdit) - it has nice, very helpful feature - preview with option (if i recall correctly) 'leave noise' - this work exactly in opposite way - remove signal leaving only noise - this allow to find sweet spot where you adjusting levels and spectral characteristic to point where only noise is audible - noise itself can't be modulated - if you can hear anything tonal (like tones similar to your signal) then it is wrong and you need adjust your settings to point where only hear able audio is only noise - you need be patient and prepared for lot of tries.

[brispuss]

try a stem splitter

Just noticed your post.

Never heard of stem splitter before. Your processed sound seems to be pretty good, thanks.

Using SVsep.com? This is an online tool I believe, and although it may produce good results, I am not keen on uploading files anywhere for processing. I would prefer to use a standalone/downloadable application on my computer.

I'll look into stem splitters in a bit more detail.

[brispuss]

Blind noise removal is not easy task - perhaps there are some "AI" novel methods working better than classical approach.
FFmpeg offers some audio denoising methods (afftdn , anlmdn , arnndn).
For neural method check https://odan.github.io/2023/08/20/reducing-background-audio-noise.html

For classic approach(spectral deconvolution with help of FFT) i use Adobe Audition (former CoolEdit) - it has nice, very helpful feature - preview with option (if i recall correctly) 'leave noise' - this work exactly in opposite way - remove signal leaving only noise - this allow to find sweet spot where you adjusting levels and spectral characteristic to point where only noise is audible - noise itself can't be modulated - if you can hear anything tonal (like tones similar to your signal) then it is wrong and you need adjust your settings to point where only hear able audio is only noise - you need be patient and prepared for lot of tries.

Noted, thanks.

I have ffmpeg, which I often use for various tasks.

[davexnet]

I tried music rebalance (step split) in RX; the results were interesting, the music and noise were
in the same stem and the little percussive sound was separated out. It may provide a bit of flexibility
to work on it this way.
Pandy's suggestion seems interesting although I think you have to install GIT and possibly Python
to build and install the models. This is new to me I'm not a programmer. On the other hand
it might be useful opportunity to learn Vapoursynth

[smartel]

Ultimate Vocal Remover (UVR5) maybe but i never tried it


https://sourceforge.net/projects/ult-vocal-remover-uvr.mirror/files/v5.6/

[davexnet]

I tried the FFmpeg noise reduction models suggested by Pandy but no luck. According to the info txt files, they're trained on voice recovery
and unfortunately not effective on this material. Looks like you're stuck with traditional methods for now

[brispuss]

Additional comments noted, thank you.

Tried UVR, but it didn't seem to do anything? Maybe I didn't use the right settings (used Denoise model)?

Also tried Steinberg SpectraLayers. This does reduce the noise to a degree, but not getting good results.

I might have found an alternative source for cleaner sound with little (or no) noise. In the process of checking this alternative sound out.

EDIT: The alternative sound track although fairly clear with little noise, it is not the same as the original soundtrack with noise. So the alternative sound track is not really usable. However, if I can't find a clear copy of the original sound track, I may have to use this alternative sound track as a last resort.

[davexnet]

UVR can split the vocal from the instrumental (as far as I can see looking at it for 10 minutes)
I don't think it's going to help clear up your audio.

Here's a sample using the MDX model.
That's a shame about the alternate source

[pandy]

I tried the FFmpeg noise reduction models suggested by Pandy but no luck. According to the info txt files, they're trained on voice recovery
and unfortunately not effective on this material. Looks like you're stuck with traditional methods for now

Try different models - this is quite crucial - personally i still prefer traditional methods - key to success is preventing spectral leak thus multiple tries required.

[davexnet]

Hi Pandy - are there any further models that will work with ffmpeg? The GIT
build put the models in \Windows\system32\ which I thought was weird
( I'm new to the GIT environment!)

[davexnet]

I found a site with a good DeNoise AI model, worked well on your sample:
https://mvsep.com/en/home

[brispuss]

Thanks for your post.

As mentioned in my post #9, I am not keen on uploading any files to online sites for processing. Regardless of how well online sites processes files, I don't know what those sites will do with your files after uploading and processing. I prefer downloadable software installed on my computer for processing on my computer only.

Using Izotope and it's Spectral De-noise model, I am currently in the process of removing or at least minimizing noise by using the "learn" function of Spectral De-noise, as I have found very brief noise only section(s) of the entire audio clip for Spectral De-noise to 'learn'.

But I've noticed the noise in different parts of the audio clip doesn't seem to be the quite same. So noise learned by Spectral De-noise for one part of the audio clip may not work well at removing noise from other parts of the audio clip. This means that Spectral De-noise will have to 'learn' from other parts of the audio clip for the other noisy parts on a case by case basis. However, the difficulty is in finding noise only parts for Spectral De-noise to 'learn', and this at the moment is proving to be difficult because other noise parts seem to contain wanted music, voice, or effects.

I'll keep searching for noise only parts, even if the noise only parts last for only a second or so!?

[davexnet]

That's fair, I had forgotten your preference regarding the uploading of files.

What version of RX are you using? If you're using a recent version, it has the "adaptive mode" that calculates and adjusts its
noise profile as it operates. I couldn't tell if would help because the sample is very short.

It's interesting that the AI divided the file into noise and tonal, similar to Deconstruct - only the RX tool
isn't capable of that level of separation.

Good luck!