Techniques for recovering poor audio from video recording

[TounLU]

Dear SOS Community,

I do work almost daily on audio dialogue which is not perfect and I try to improve the way it sounds.
I use multiple tools with more or less effect, it really mostly depends on the quality of the recording on how good the result is. (Adobe audition/waves plugins(clarity)

Now I received a poor audio from a video, which features multiple issues and I wanted to ask if you might have a good way to approach it.

Now typically I would normalize the audio, manually lower excessive peaks and normalize again, try to remove constant background noise.
Next I would use a deverb to get the room noise under control.

Typically this works quite well, later I push it into a vocal chain where I EQ and Compress and limit to requested target.

Now because this audio is recorded way far away from the mic (my guess is the onboard mic), the deverb plugins don't give me a good result.

My question is, might it be better to first compress and EQ or use another approach to the sound for control first so that the denoise and deverb work better?

What are your experiences on this that you could share?

I feel my result sounded way to robotic, I lose a lot of warmth in the sound of the voice, so I really wonder if there are ways to maintain the sound yet clean it up.

I would provide the audio but I am not able because of rights.

[James Perrett]

The first thing I would suggest is to do the background noise removal as the very first thing. If you make any level changes before removing the background noise, you may confuse the noise removal algorithm.

I would also suggest using Izotope RX instead of Audition for noise removal. In my experience RX is much better at noise removal than Audition (and anything else I've tried) although Audition still wins out for some other processes.

(NB - I use an old version of Audition before it went subscription only so things may have changed).

[Wonks]

If you are part of a team making videos, then getting them to record the sound better at source is obviously a lot easier than trying to fix it afterwards.

[Tim Gillett]

Hi TounLU, advice often needs to be tailored to the actual recording. You said there are rights issues with the audio but a short audio file exerpt, even if only of the noise when nobody is speaking would at least illustrate the nature of the noise.

[Matt Houghton]

As others have said, I'd start with any constant noise. I tend to use RX's spectral denoiser for that. Sometimes it works better in two passes than one. What I did next would depend on the desired/acceptable quality for this video, whether it's just dialogue, if so whether you need to preserve the ambience at all, and how much time you have to chase the best results.

But for a quick fix-all sort of thing for dialogue, you could give Accentize dxRevive a go. There's a free demo. Just listen out for any artefacts later if you're compressing, de-essing, EQing etc. I quite like that this is an insert plug-in because you can audition the effect of such processing before you commit to the settings.

[PippaPumpkin]

Hello,

I am a hobbist and with my little experience I agree with what has been said here. If you can get the production team to up their recording skills, that would be the best long term solution, as Wonks has said.

I always denoise my audio first thing. This is especially helpful for me when there is a consistent noise floor. I use a freeware software for now, but I would recommend going for something like RX since the freeware software I use is no longer supported and can cause crashes.

If you can't provide a sample of the audio you are having trouble with, maybe you can find an example that illustrates the problems your having.

I hope you can make it work.

[TounLU]

The first thing I would suggest is to do the background noise removal as the very first thing. If you make any level changes before removing the background noise, you may confuse the noise removal algorithm.

I would also suggest using Izotope RX instead of Audition for noise removal. In my experience RX is much better at noise removal than Audition (and anything else I've tried) although Audition still wins out for some other processes.

(NB - I use an old version of Audition before it went subscription only so things may have changed).

Would it not make sense to like max out the file by normalizing it, all you really do is raise the highest peak. In my opinion from what I heard, some effects tend to work better this way. But I get the point you make. Assuming the noise algorithm works this way which is hard to justify without knowing it

I am aware I might not use the best tool, but this is what I am provided with from the corporation I work with and it is what I have to use. I do have RX7 elements. I used it for a while but I did not update and I phased it out of my workflow because I can work quicker in audition because I constantly had to move files from one place to another.

I know there are also other post options like spectralayers which I saw in the mag a while ago, looked also like a good option to try one day.

[TounLU]

If you are part of a team making videos, then getting them to record the sound better at source is obviously a lot easier than trying to fix it afterwards.

I am fully onboard with you on this, but inside a huge corporation people you have nothing to do with sometimes need your help after they messed it up.
Trying to teach people feels like an endless task sometimes. There are always new ones who think it all works automatic just like that

I just speak and I sound like a superstar in some A level production

[Wonks]

It's good to know the background context.

So, apart from trying to educate people, you are pretty much stuck with fixing things.

[Drew Stephenson]

I am fully onboard with you on this, but inside a huge corporation people you have nothing to do with sometimes need your help after they messed it up.

Can relate!

[Bob Bickerton]

Just to add to what Matt said, I’ve recently discovered the plug ins from Accentize and found that DxRevive is pretty amazing at getting a good result on much material without the faffing of RX.

Well worth trying out their products: https://www.accentize.com/

Bob

[Tim Gillett]

A recurring problem with such products including Accentize can be their choosing speech samples which are already perfectly intelligible, and even with more of the same background sounds would still be perfectly intelligible. In other words the original samples seem to have been carefully chosen to flatter the tool and conceal its limitations. Purchasers working with much more challenging audio examples can start to believe they are doing something fundamentally wrong. But the real problem can be the deceptive advertising. It's not so much the audio examples the seller uses but the difficult examples they dont use.

[Matt Houghton]

I've used a range of such tools on a range of material, Tim, the most notable exception being CEDAR's stuff. dxRevive isn't perfect but it is a lot better than most — currently the o e I'd choose over others. I've recently used it to rescue some pretty ropey dialogue recordings — built in computer mics in horribly reflective rooms, with background noises. If there's no prospect of rerecording, it's well worth a try.

No polishing can make a turd appealing, but I do reckon this could easily be the difference between a recording being usable or not.

Another option might be Descript's Studio Sound, which is an online service. But you have less control with that, just a wet/dry % setting.

[BWC]

Would it not make sense to like max out the file by normalizing it, all you really do is raise the highest peak.

Doesn't make sense to me. Plus, you said that after normalizing you "manually lower excessive peaks and normalize again." That's definitely changing the audio (including the noise), very possibly enough to confuse a noise removal algorithm.

[Tim Gillett]

Matt part of the problem as I see it is the absence of a set of industry standard audio test examples with a range of levels of difficulty, from easy to impossible. In the early days of CDs I think Philips provided for service techs a standardised damaged CD. How well the CD player's error correction negotiated the CD could be easily tested and heard. That sort of thing.Without that buyers without the necessary testing expertise can be left floundering.
On the other hand, in the "learn" section on their website Izotope offers tips as to which order of tool processing is best for audio restoration but not all practitioners follow this apparently good expert advice. Eg: some advocate using a general denoiser tool as the first restoration processing step, whereas Izotope recommends using it as the last. Almost a reversal. For those trying to adopt the best techniques this can be confusing.

[Bob Bickerton]

Tim, I’ve just finished a project processing multiple and old video audio clips, many taken on basic camcorders, no bespoke microphones, in highly reflective halls. I found the Accentize plug ins gave me as good results in much less time than RX, I think better in some cases.

Perhaps for the benefit of readers here you could clarify if you’ve actually used dxRevive? Your answer will give efficacy to your criticism …….. or not.

Bob

[Tim Gillett]

If you are part of a team making videos, then getting them to record the sound better at source is obviously a lot easier than trying to fix it afterwards.

That large slice of common sense wholeheartedly seconded here! Well said Wonks.
Audio processing material already recorded sometimes long ago, and which cannot be reperformed and recorded is a significantly different situation, obviously.

[Matt Houghton]

Matt part of the problem as I see it is the absence of a set of industry standard audio test examples with a range of levels of difficulty, from easy to impossible. In the early days of CDs I think Philips provided for service techs a standardised damaged CD. How well the CD player's error correction negotiated the CD could be easily tested and heard. That sort of thing.Without that buyers without the necessary testing expertise can be left floundering.

A problem with that approach, Tim, is that, right the way from the easy to the impossible, there can be myriad different problems and combinations of problems to contend with. It just wouldn't be feasible to create such test files. Another problem is that what is deemed right / wrong / good / acceptable / unacceptable in terms of dialogue restoration is very much in the ear of the beholder — for CD error correction there's very definitely a perfect target to aim at. So I don't think that's a great analogy. Fortunately, most of us who edit a range of varying quality dialogue regularly have huge amounts of material available to test and compare these tools...

I agree with you on using RX. Learning how to use it (and various other tools, such as Acon's and Accentize's) to best effect takes time and experimentation. iZotope have their own guidance, of course, but even they haven't always hit on the best way to tackle every problem. (Indeed, often, I find that a tool in RX that isn't the one designed for the job can be more effective that the one designed to do the job.)

But seriously, dxRevive is very different from RX. And they can both be deployed on the same source to good effect. If you've not tried it, you shoudl give it a go.

[Tomás Mulcahy]

Would it not make sense to like max out the file by normalizing it

No it does not make sense. James is correct. The less you do to the file before denoising, the better. Premier has had a very good denoise plugin for a few years now, it is AI so you don't really need to buy anything else. It gives you limited controls to focus on the noisiest area of the spectrum, and a percentage reduction slider. It is specifically designed for badly recorded stuff! Use it in real time via the Audio Track Mixer, not the clip based Effects pane.

Adobe have this as well, it's a different process still AI:
https://podcast.adobe.com/enhance

[Tim Gillett]

Matt part of the problem as I see it is the absence of a set of industry standard audio test examples with a range of levels of difficulty, from easy to impossible. In the early days of CDs I think Philips provided for service techs a standardised damaged CD. How well the CD player's error correction negotiated the CD could be easily tested and heard. That sort of thing.Without that buyers without the necessary testing expertise can be left floundering.

A problem with that approach, Tim, is that, right the way from the easy to the impossible, there can be myriad different problems and combinations of problems to contend with. It just wouldn't be feasible to create such test files. Another problem is that what is deemed right / wrong / good / acceptable / unacceptable in terms of dialogue restoration is very much in the ear of the beholder — for CD error correction there's very definitely a perfect target to aim at. So I don't think that's a great analogy. Fortunately, most of us who edit a range of varying quality dialogue regularly have huge amounts of material available to test and compare these tools...

I agree with you on using RX. Learning how to use it (and various other tools, such as Acon's and Accentize's) to best effect takes time and experimentation. iZotope have their own guidance, of course, but even they haven't always hit on the best way to tackle every problem. (Indeed, often, I find that a tool in RX that isn't the one designed for the job can be more effective that the one designed to do the job.)

But seriously, dxRevive is very different from RX. And they can both be deployed on the same source to good effect. If you've not tried it, you shoudl give it a go.

Matt, perhaps I wasnt clear enough, hence the following:

My experience is that on their websites, vendors of audio restoration software, but particulary of broadband denoising tools, tend to choose their unprocessed sample files so as to best present their tool's effectiveness and - perhaps more to the point - disguise its limitations. They give the tools relatively easy test examples, making the tool really shine. Unless the potential purchaser is aware of this, they may buy the tool, believing (from listening to the vendor's curated samples) the tool can excel at much harder cases.

That's why I'm all for vendors offering free trial software! Then at least the potential buyer can test out the tool on much more demanding samples than presented and decide before they buy.

But that's also why I'm still hopeful vendors will demonstrate their software using more demanding audio samples, designed by others with no skin in the game and with no perverse incentives at play. Hence standardised tests for particular tools.

It's not rocket science. I made up such a sample as a training exercise 13 years ago. It was a simple mix of a voice with heavy white noise. I lost it so I recently made another one. It really does expose the limitations of a general denoiser tool even today, just as it did 13 years ago. The sample was uploaded to GS but interestingly not one reader in an audio for video post production sub forum was prepared to try their choice of software on denoising it and upload the results, even though GS allows easy direct upload of audio files to the discussion threads.

So as I've been doing at times, I'd encourage the OP to take advantage of free restoration software offers and especially to test with difficult audio cases, much more difficult that the vendors seem to demonstrate on their websites. The only way to find the limits of the tool is to test it on samples which do just that.

[Matt Houghton]

Well, most *do* offer a free demo. And adding white noise to speech isn't that good a test for some of these things — particularly the ones that are supposed to remove any noise that isn't dialogue, including artefacts of lossy compression, or the ones that attempt to synthesise LF/HF that has been lost due to filtering somewhere along the way, or to remove/reduce 'room tone'.

The whole point of most of those videos is to showcase what a product *can* do. A marketing tool to grab your attention. An advert, in other words. Something to tempt you to test the demo, and then make up your own mind...

That said, I've seen a lot more manufacturers (not of denoising software, I hasten to add) recently accompany such material with a longer 'deep dive' video, in which they'll use the product on a wider range of material and comment on what's working/what isn't, or what tricks/other tools might help you get the best results. I'm all for that!

[Tim Gillett]

...Now because this audio is recorded way far away from the mic (my guess is the onboard mic), the deverb plugins don't give me a good result...

I've listened to the six Accentize dxRevive website 'before and after' samples and probably unlike your problem video sound tracks, all six untreated samples are perfectly intelligible as they are with no processing. (Unfortunately this seems quite common in vendor 'before and after' audio examples) The original tracks have only moderate reverberation. All seem to have been recorded with a relatively close mic, or perhaps outdoors where reverb from using a distant mic isnt normally so much of an issue.

So whether dxRevive will help with your recordings of voices "far away from the mic", only you can know that by trying out the software.

[TounLU]

First off all, thanks for your many replies here!
I was locked out of the forum because of some bug and was not able to see anything.
Thankfully support helped me out, but I was so busy the last days I did not have time to make a proper response.

Would it not make sense to like max out the file by normalizing it, all you really do is raise the highest peak.

Doesn't make sense to me. Plus, you said that after normalizing you "manually lower excessive peaks and normalize again." That's definitely changing the audio (including the noise), very possibly enough to confuse a noise removal algorithm.

Indeed you raise a good point, I will try this next time. If the algorithm is trained on certain noise patterns changing the dynamics certainly can throw it off.

Just to add to what Matt said, I’ve recently discovered the plug ins from Accentize and found that DxRevive is pretty amazing at getting a good result on much material without the faffing of RX.

Well worth trying out their products: https://www.accentize.com/

Bob

I did catch an article in the current SOS about one of their products. It certainly looks very interesting and worth a try! I also see other people mentioned it here and there has been good points raised.

Would it not make sense to like max out the file by normalizing it

No it does not make sense. James is correct. The less you do to the file before denoising, the better. Premier has had a very good denoise plugin for a few years now, it is AI so you don't really need to buy anything else. It gives you limited controls to focus on the noisiest area of the spectrum, and a percentage reduction slider. It is specifically designed for badly recorded stuff! Use it in real time via the Audio Track Mixer, not the clip based Effects pane.

Adobe have this as well, it's a different process still AI:
https://podcast.adobe.com/enhance

Another good option in case I do not have access to other tools. Thanks!