izotope RX is really good better than wavesIMHO but when you get into the bigger leagues cedr and cubetec are incredible....just some of the feature in them are mind blowing.

Branko is absolutely right, as usual. Your samples demonstrate a "brute" usage of broadband noise reduction for something that is cyclical. For a complex "buzz" such as in the example, I would not first reach for any of the things that you "demonstrated". Furthermore, when it was done, the settings were all wrong and you cut way too much. I have at least half of those and can tell that the operator was expecting a "different" result and thus cranked down on the NR Amount. They all sound muffled and over-processed.

The first thing to do in a case like this is to start with EQ filters. For a Voice, we don't need anything below 80-90-110 (depending) so start with a hi-pass. When you say that your practice has shown that denoisers work better than notch filters I wonder if you are only using one notch. A set of harmonic notches is the thing to use in this case. Start with that BEFORE you put it into a Broad Band denoise filter. Once you get good buzz reduction with properly placed harmonic notches, then use a de-buzz-ing plug in (unless it is a simple notch set). THEN move to the broad band and use much more judiciously.

I messed around for about 15 minutes using a combination of techniques and acheived a more acceptable, less damaging to the voice result than the "survey".

The Comparison used the wrong tool for the job. You have to learn the different types of noise and how to attack them.

I find Izotope useful for some things and useless on others. The De-clipper is jaw-dropping WHEN IT WORKS. I have thrown many things at it that it COULD NOT fix. I find Wavearts useful on some things, but I hear it's artifacts very very early in the process so it is not my favorite. NoNoise has a clunky workflow, but can excel on many things (despite it's very very old DSP); on other things it can't help. Sony Oxford Denoiose is pretty useless EXCEPT for the DeBuzz.

The Orange Box is STILL quite useful and clean and artifact-free. You simply go too far with it and back off. (Thanks TVPS, and I think, Marti). But I doubt I would reach for it here, or if I did, only "tickle" it.

Cutting 12 dB from broadband noise and at least 12 dB from the hum is the test condition. If you can do it better with these tools - why don't you post the result? Using noise reduction for "cyclical" noises is not a "brute" force, it's one of intended uses of noise reduction. This is emphasized in the manual of TC BackDrop or iZotope RX, for example. So, you are just not familiar with this use.

Of course not. I've been using a set of notch filters, available in DeHum modules of many software.

I really know different tools One of my goals has been to demonstrate that broadband noise reduction can succeed on tonal noises where notch filtering fails. Notch filtering mostly fails on this sample.

Didn't see where it said the criteria was -12dB of noise reduction using this Plug-in only. I only was the original files and the treated ones with comments. Now I know what the test was testing.

However, this is not how I would approach this noise problem.
This was done quickly, with almost exclusively notches. I can still hear the buzz, but it is faint. With a little more futzing, I think I could get rid of the rest.

This amount of suppression is quite sufficient, actually you suppressed it more than required. However everything comes at the price: in your result the timbre of voice sounds distorted to me - dull and echoed. Here are the results that I could achieve with noise reduction, w/o any notches.
Speech_Test0.wav is just hum reduction.
Speech_Test1.wav - with around 12 dB of broadband noise reduction as well.
Try switching between the original, these files, and your result. You'll hear what I'm talking about: the timbre of speech is retained much better in my results. And this is because no notches have been used.

Here's a quick attempt. Two files: one of them with notches only (Equium) and other with notches and NR (Equium and Cedar DNS 2000 - identical to DNS 1000, but with software controller)
Notice how breaths are not affected.
It could be a little better, but I'm in another studio, not mine.
I'll be back to my studio next week, and try to denoise with MR Noise, Sonic Solutions NoNoise, etc.
Branko

Hi Lukin, Branko et al,

I would like to point to Wave Arts's Multidynamic 5.
Minister, you have mentioned the Orange Box (Dolby Cat 43).
Well, Multidynamic 5 has the same philosophy than Orange Box but instead of 4 band you can have till 7 bands.
I got very good results with this plugin.
Just to mention it.

And the winner is...

Branko! (Equium + DNS)

IMHO

Yes Tom, it was me. The CAT 43 is a great box. I haven't used one in years though.
I haven't heard the above tests yet but I'll throw my comments in a little later.

Naw, it was me, Tom bought my orange box!!!

TVPS,
I didn't say I sold it to him, I just told him about it.
This was years ago....

Hi Guys,

No time today. It is 11:30p and I am just knocking off having mixed a noisy TV show and dealt with a new employee... My ears are cauliflower.

Alexy, as I said, I did it very quickly, but, yes, I went too far. If I can find time in the next day or so, I will open the file again and backoff on what I did. There is distortion in the recording, but, ironically, the little bit of distortion in mine is from SONIC that I put at the end.

More on that another day.
Hi Jose,

Yes, thank you, I have been using the WaveArts for over 2 years now. In 2 different rooms. It is the sample principle as Cat 43/360, but it is different and much easier to get into pumping and arifacts with the WA. I use both.

The WAVES C4 will also do this.

Yes, Marti told me about the 43 years ago (like I actually listen to what he has to say over the phone...I like to at least humor him.), and yes, I now have TVPS's and he can't have it back !! But Marti I thought you suggested he send it to me, but this was after we made the deal !! So thanks to both of you. Now can't we just get along?

Thanks for making the samples. I have to point out one inaccuracy that prohibits fair comparison: your levels are dropped by almost 6 dB and the sampling rate is different. Here's what you get when you normalize the samples to match in level (see attachments).
Also, my own opinion is that my result of hum removal sounds better than yours - there's less damage to the voice.
On broadband noise reduction - I agree that in your sample breathing sounds have less "musical noise" artifact, but at the same time they are significantly suppressed in level, and also have noise modulation.

Thanks for the gain-bumped versions. I didn't have the time to do it myself yesterday.

For me the winner is Isotope RX version 1 (Speech_Extreme.wav) if I would have to use the clips as-is without other sound elements (like AMB).

If this scene had a TV set (телевизор) that is yelling (кричать) or other noisy source, I'd use Branko's Equium+DNS version, because the voice is very slightly more natural, but gate artifacts on breaths would have to be edited.

On the other hand, Speech_Extreme.wav has the least artifacts overall, and I would like to see a version with reduction backed up a few dB, with a band-gate (like C4) added to suppress some noise between the words.

To me, this test has truly revealed that a broadband reduction plugin can give results very similar to notches + NR, while it is probably much quicker to implement. Off course, we don't know the exact inner workings of Izotope RX, so it may have some sort of intelligent notches built-in, because it excels in hum removal over other broadband NRs.

It has separate controls for broadband noise and tonal noise and can remove them by different amounts. What is more important - it doesn't suppress the useful signal at frequencies of the hum when the signal is well over the amplitude of hum.

sorry, I was listening on a dub stage, and the level was just too high, I forgot to bump it up before recording... dub stage is an expensive tool, every minute counts.
Analyzing the results and listening on headphones, I can hear a phasey component caused by multiple notches in my samples. This is due to the lack of time to fine tune the filter set and to clean the overlapping bands (I was just adding notches one by one, without reviewing their efficiency), so there are too many notches applied. Every notch, by it's nature, causes some ringing, so adding tons of these will produce phasey effects.

However, the purpose of my attempt was to show that you were sometimes using tools in a wrong way, especially DNS 1000, and that is what I tried to stress in my first reply to this post.

One conclusion can be drawn from this little exchange of opinions: there are definitely two different cases where we need a noise reduction:
1. everyday work, dialog pre-mixing, where we need a fast and reasonably efficient tool that would do no big harm to our material, and
2. special cases (could we call it "restoration") where we need to apply all our knowledge and tools to achieve the best result possible.

Needless to say, I'd ask the ADR to repeat the example recording. In order to clean it properly, you'd need half an hour. That is way too expensive. Also, if there's a TV set in picture, I wouldn't clean it up that much.

Branko

p.s. as soon I get back to Lisbon, I'll post some new results. I'd also like to see a different shootout, for example a low dialog with a sea wash background, a/c stuff and so on...

I take it back. That 6db bump made a big difference.

I may now be leaning towards Speech_Test1

I can't wait to hear Branko's next take.