Noise reduction (in Adobe Audition)

[Peter Berger]

I need to reduce noise in some inteviews and also increase the volume. Its better to reduce noise first and then icrease the volume or vice versa? Maybe it doesnt matter?

[Richard Crowley]

It would probably be best to try it both ways yourself on a sample of your own recording.
Also, probably faster than waiting for an authoritative answer here (if there even is one).

[Greg Miller]

I've always boosted the level first, going on the belief that you want the NR corrections (e.g. level changes) to be as fine as possible, relative to the overall level. (I suppose, if you extrapolate from that, that it would be even better to convert to 32 bit, but I've never carried things that far.)

I've also found (embarassingly, after I had been using the program for years) that NR seems to work much better if you do it at 96kHz sampling rate... even if the original was 44.1 or 48.

I don't have any quantitative measurements or heavy math to support the above... just my qualitative evaluation when listening to the results. YMMV.

[Rick Reineke]

Never used the NR in Audition but I usually use SCS NR-2.0 as the first process in a chain when needed. If significant NR is desired, two instances NR processing will produce fewer artifacts than one with a high reduction parameter setting, especially with an updated 'noise print' on the second.
Interesting theory Greg, it would have never occurred to me to use a higher sample rate. I'll have to try that. Thanks

[Greg Miller]

Rick,

I consider that every digital sample is an inaccurate representation of the original analog signal. There is some error between the original and the digitized value. The size of the errors can be reduced by (1.) digitizing at the highest reasonable level, (2.) digitizing with the greatest reasonable sample depth, and (3.) digitizing at the highest reasonable sample rate.

So I try to give the NR algorithm the greatest possible number of bits to work with, within reason. I'm pretty sure that Audition does all internal calculations at 32 bit, so I didn't bother converting the file to that depth. But 96Ksamples/sec did seem to make a difference. That might or might not be noticeable, depending on exactly what tool you're using.

BTW, I agree wholeheartedly: two passes of light NR, with an intermediate noise sample after the first pass, turns out better than one heavy-handed pass. And I try to do the static things first, then the dynamic things last. e.g. if there's a heavy 60Hz hum, I try to remove at least some of that with a fixed notch filter first. Then I take a noise sample and run the dynamic NR second. And if the result isn't quite as quiet as I'd like, I might run a very slight bit of downward expansion as the last pass.

In reality, I should probably be saving these intermediate steps as 32-bit files, because the accumulated errors would be less than when I save each step in 16-bit form. Sigh.....

[Rick Reineke]

I do about the same workflow, HP, LP filtering prior to NR I usually work in Sound Forge Pro and Vegas Pro which use 32 floating bit point.processing. I'm also experimenting with SpectraLayers Pro but the learning curve is steep . I would think that is 32 bit like other SCS apps.
So in my case, do you think re-sampling to 96k, process, and then down-sample back to 48 would not be of much benefit?

[Greg Miller]

I can't really give a definitive answer to that. It's probably a function of the internal math, and that's going to vary from program to program, and from filter to filter.

I only know that I've gotten better results when I upsampled to 96k before doing the NR. But I suppose the only way to answer that for you, with your software and your material, is to try it. (I feel fairly safe in saying that working at 96k can not make the result any worse.)

And, as I said, I should probably convert the file to 32bits, and save all the intermediate steps at 32 bits, then downsample to 16 bits as the final step. That's because, even though the processing is at 32 bits, if you then save that step as 16 bits you're introducing another "rounding error" in the math. As you have more and more processing steps, and more and more intermediate files, those "rounding errors" will accumulate. I would think this applies to any program and any process.

As I confessed, I haven't done that myself. When I've been doing the really extreme cleaning, it has been on some old material which would never sound really new or pristine, so I thought saving 16-bit files would be adequate. But in theory, and with newer better source material, yeah, I'd probably save everything at 32 bits.

(What's the analogy with images? Wouldn't you do all your work at the highest resolution possible, and then if you need a low-res product, do your final downsizing as the very last step?)

[Rick Reineke]

Myself and most other pro production sound mixers record @ 48/24. No need for dialog @ 96k. My 744 and other location recorders don't offer 32bit as an option anyway.
No to get to far OT, but for video I always render directly from the HD camera files. except for material destined for YouTube, Vimeo or other web streaming, Then it is rendered to an intermediary Avid DNxHD format, then re-render to MP4/H264 in 'Handbrake' ... which seems to be the preferred method for many skilled video folks.. who know a hell of a lot more me, The final picture quality is impressive. This tutorial is from my friends over at the SCS Vegas forum.

Sony Vegas to Youtube Tutorial - A Better Method - YouTube

[Steven Digges]

Another great thread with solid advise. I am going to try the multi step process myself next time I need to. Thanks guys.

To the OP. I am not a great audio guy but I can work my way around the fundamentals. What I would say to you is take everything in small steps. Experiment a lot. Sampling your “bad frequencies in Audition and zaping them out of the track will sound terrible if overdone. I have learned good finished audio is a lot like ones taste in food, we are all different. I have worked with one sound guy who knows far more than me about the advanced tools. By the time he is done “sweetening” I never like the result. To him it sounds good, to me it does not sound like the voices we interviewed. I try to do as little as possible with the goal of replicating the real sound in mind.

Steve

[Rick Reineke]

"To him it sounds good, to me it does not sound like the voices we interviewed."
> If you can post a before and after clip, we ('pro' audio folks on this forum) can evaluate your 'sound person's' work. Maybe just a want-to-be with some audio tools. Lots of those types around.

[Bob Hart]

When I had have to do some serious noise reduction in Cool Edit Pro, Adobe Audition's ancestor, my personal preference has been as Greg Miller suggests, operate in the 96Khz environment :-


Normalise to 95%.

Take a noise sample from a dialogue-free portion of the recording.

Apply the noise reduction to a level where there is reduction but no digital artifacts like tinking or bubbling.

Sometimes I will add a different constant-level background noise of a lower frequency than voices in the hope of masking more extreme noise reduction artifacts if I feel I have to push the noise-reduction too far, a sort of a white noise thing. Often does not work.

If I have two tracks, either stereo or mono copied to a second track, I sometimes invert one of the tracks to see what happens. Sometimes you strike it lucky but there is no science or logic to this.

I use the graphic equaliser in 30 channel to see if I can notch out something undesirable manually in the quieter patches between voices using the zone selection ( highlight ) function.

I try a little compression only. Too much and the noise may "pump".

I select the quiet pieces between the dialogue and lower the audio level. I do this before adding an artificial background or ambience track in a sometime futile effort to mask my errors and any compression pumping.

I try delay and play with the manual settings. Sometimes something annoying may cancel out if you have a mono track, copy it into a second track and apply the delay effects.

I am no expert in this at all so please take notice of other comments over mine.


This old damaged acoustic 78rpm recording represented one of the most extreme challenges I have faced. It remains pretty bad.

YEARNING - GERALD SCOTT.mp4 - YouTube

The surface damage and wear damage in the groove was rather severe.

There is a mid-frequency artifact which little could be done about. The substrate of the old shellac surfaced disks is reinforced with a woven binder which seems to have the same texture and pattern of old hessian sugarbag material. This pattern is apparent in the surface of the shiny smooth shellac layers which have been glued onto the substrate.

It reproduces as a cyclic and varyingly random low to mid-frequency rise and fall which alters as the stylus migrates to the centre of the record.


This one was less of a challenge but I found it necessary to selectively draw in with the graphic equaliser some of the critical consonents of the words where I was able to and lower levels between some verses to eliminate compression pumping. You will hear where I went a bit too far with delay effects with a reverb on a few strong bits.

BOB DALE MARY'S BOY CHILD - YouTube

I had to draw out a few bad pops manually by stretching the waveform and re-levelling a highlighted very small time period. These digital audio tools are indeed wonderful for damaged recordings. They also enable one to replicate the original equalisation curves the engineers intended before the common RIAA standard was introduced. I have found constant background noise in video recordings a much harder beast to deal with. .

[Nicholas de Kock]

Adaptive Noise Reduction in Adobe Audition is pretty impressive when working with voice. I usually first run Adaptive Noise Reduction then increase the volume.

[Greg Miller]

Originally Posted by Bob Hart

If I have two tracks, either stereo or mono copied to a second track...

Bob, when I'm cleaning an old mono phono recording, I always record it in stereo (after all, the phono cartridge and preamp are stereo anyway). First I de-click it. Then I listen, comparing channels. Sometimes there is a bad pop/click remaining on one channel only, and I can then select audio from the cleaner channel.

Ultimately, if both channels have similar S/N ratio, I sum both channels to mono, which usually makes the noise level sound lower.

Originally Posted by Bob Hart

I use the graphic equaliser in 30 channel to see if I can notch out something undesirable manually in the quieter patches between voices using the zone selection ( highlight ) function.

For noise with a steady fixed frequency, I use the parametric instead. First I display the frequency spectrum and get a rough idea where the noise frequencies are located. Let's say I find noise approximately 100 Hz.

Then I open the parametric. Select one filter only. Set Q very high, usually = 50. Set frequency somewhat lower than where the noise is... in this example around 90 Hz. Drag the gain UP to somewhere around +12 to +20. Click "Preview" to start the audio sample playing. Now start increasing the filter frequency, and listen for loudest output. (In this example, it will be somewhere around 100 Hz, but maybe a little higher or lower.) Sweep it up and down until you have the frequency accurately identified. Once you find the frequency with the loudest output, you know your filter is tuned exactly to the noise frequency. Now pull the slider down into the -dB range, and you will hear the noise go away. Don't pull it down more than necessary. Finally, repeat with other offending frequencies.

Sometimes I find this gives better results than the adaptive noise filter, because the parametric doesn't produce any "burbly" and "tinkly" artifacts. You have to use it with caution if you use it within a part of the file containing desired program material, because you will also notch out pgm. matl. with the same frequency.

[Al Bergstein]

Great thread, thanks for this insight. I've been also struggling through figuring out the proper process and have not thought about upsampling first. I'll save this one.

[Bob Hart]

Greg. - "Record a mono track in stereo". - Yes. I also do this. I have an old AWA gram which has a "mono" setting which somehow seems to clean up a little on its own. The waveform off it is exceptionally weird with the peaks on one side of the centreline about twice as deep as their opposites.

For this and a later electronically controlled turntable, I ended up creating a filter I called de-RIAA and then created other filters for each of the equalisation charts I could find for the various record labels of the times. For what they were, there is a surprising amount of information in those old recordings.

I don't remember seeing "adaptive" noise reduction. It may have been there under another name or function. My latest was Cool Edit Pro v1.5. Parametric equaliser. I shall have to go there. I am not sure if it was in CEP. It may have been that I simply did not dig deep enough. I did use parametric equaliser on a Fostex DMT8 in earlier 78rpm recoveries. To get a full re-requalisation I had to bounce-track it one generation. I was also using and please don't laugh, the dynamic effects of an EM-U ESI-32 sampler with 64Mb of memory for some cleanup. It did quite well but one 3 minute track would take all night to run through and one hoped one got it right.