Hi Paul, nice to meet you too. I see you're located in "London-ish". So am I (Watford, actually). So if you feel the urge to meet up sometime it could probably be arranged. I once met Derek Higgins (author of Wave Corrector) and we had a good old chin-wag.

I agree about the need to zoom in [in spectral view] to see the clicks. It's the same with Wave Repair. My personal view is that spectral view is useful when you're doing a manual repair by listening. When you hear a click, zoom in around the general area, switch to spectral view, and the click stands out quite clearly.

Other glitches that lack high frequencies (plops and thuds) are much harder to spot. Sometimes you'll see an isolated "splodge" in the spectrogram. Quite often it's easier to spot plops and thuds in waveform view: there can sometimes be an obvious DC-like shift in the general waveform. But even once you've found such a glitch, fixing it can be maddenly difficult. I admit to sometimes giving up and leaving the original thud present, because all attempted repairs make things no better or worse.

Regarding clicks introduced by the hiss filter. This of course is pure speculation on my part, but assuming VinylStudio uses FFTs to implement its hiss filter, my experience has been that unless you apply some sort of windowing so that the overlapped FFTs are "faded in" to each other, discontinuities at the boundaries occur, leading to "ticking". A filter might remove those discontinuites. But if this is what's happening in VinylStudio, then the rumble filter is simply masking something that probably ought to be dealt with by windowing the overlapped FFTs. But I reiterate that I have no idea how VinylStudio's code actually works, so if you already window your FFTs, please don't take offense!

Seems like we're aiming at different user types. Wave Repair is primarily aimed at manual repair - fiddling with the waveform at ultra-fine levels of detail. My view is that automatic declickers sometimes do a good job on "medium level" clicks, but I've yet to find one that deals with big pops and splats, and most of them miss the tiny ticks. I'm kind of obsessive about my restorations, so I have to get in there and fix up things that auto declickers get wrong.

I've never seen VinylStudio, but from what I've read here it sounds as if it's targeted at a more automated restoration process. So the two tools could be complementary. (Seems I should download an evaluation copy and try it out). On the other hand, people who are only interested in automatic cleanup can just ignore Wave Repair - it's definitely a tool for "geeks".

CE2000 is indeed a wonderful program - IMHO the finest affordable general purpose audio editor there ever was. What a shame that when Adobe bought out Syntrillium, they chose to dump CE2000. You & I aren't trying to replace it, but somebody needs to. For sure, the install files are readily available (from my own website, for example), but activation keys are not. New users are either forced to use illegal keys or miss out on a wonderful program.

However you may also think of providing VS with an interface for “normal” users and another one for perfectionists 

I thought of applying different filtering processes to files of pure white noise (generated with Audacity). As NR is level dependent I tried also with different intensity levels: 0 dB, -20dB and –24 dB. I made these experiments with WR as it offers configurable filters. The results are interesting I think, and are as follows.

1) Rumble is better attenuated by appropriate notch filters, rather than one or two cascaded HP filters!
In my case I have rumble distributed around two peaks at 15 Hz and 25 Hz, while hum has a peak at 100 Hz and a negligible one at 200Hz.
For hum a notch filter at 100 Hz is mandatory, for rumble I experimented with one and two cascaded HP filter with corner ranging from 30 to 45 Hz.
Alternatively, I used two “broad” notch filters, tuned at 15 and 25 Hz
I found that this way the resulting response is steeper, thus frequencies between 40 and 100 Hz are less affected.

2) Applying HR filters may introduce some clipping here and there.

3) Filtering with NR is always level dependent, but if the “level dependent” option in WR is selected this characteristic is enhanced, so that NR filtering has a really negligible effect at a 0dB level,  it is nevertheless modest otherwise (without checking the option).

4) At –20dB level things are more interesting. The results, without checking the “level dependent” option, are very similar to those produced by the HR filters, but the response is steeper! Morever all information in the noiseprint are used, so that some other frequencies are lightly attenuated.

5) At the –24dB level attenuation of frequencies were noise peaks are located is greater.

6) Would it be better to apply HR filtering and then NR with a noiseprint of the residual noise?
I tried also this. The result is a slightly stronger attenuation of rumble, but at the expense of a less steep response, resulting in more degradation of (ultra)bass notes.

The only thing I could not simulate is the dynamic adaptation of the NR filter. I suppose this could be only judged by actual aural test.

However you should agree with me that I didn’t find any evidence that applying HR filters rather than direct NR is better!

You've clearly put a lot of work into this. Are you able / willing / want to post any screenshots? Upload them to imageshack or somesuch? It would be interesting to understand your tests a bit more clearly.

I'm not sure I quite followed all of that, but no, I don't agree. That soft bass note definitely disappeared, for reasons already outlined, and high levels of rumble definitely upset our hiss filter (and, I suspect, others too, we don't really do anything unusual). I still believe that NR is most effective (or, if you like, least destructive) when asked to do as little as possible. I personally would not use it in the way you advocate.

I just checked our hum filter: notches at 50 (or 60, etc), 100, 200 and 300Hz, bandwidth 0.05 octave in all cases. Does that sound right to you Clive? It's all a bit seat-of-the-pants but it doesn't seem to affect the sound.

Don't let the numbers put you off. Most of these 'false positives' do not actually matter. If you zoom in on a few of these corrections and take a look at the change that has actually been made to the waveform you will see that, in the majority of cases, it is just a tiny adjustment and is not audible. But, occasionally, such a'repair' can dull the sound of a percussion burst...

Couldn't agree more.  Adobe wrecked it.  I am lucky to have a copy.

Cheers and do PM me - Paul.

This reminds me of this other statement of yours:

Quote

Don't let the numbers put you off. Most of these 'false positives' do not actually matter. If you zoom in on a few of these corrections and take a look at the change that has actually been made to the waveform you will see that, in the majority of cases, it is just a tiny adjustment and is not audible. But, occasionally, such a'repair' can dull the sound of a percussion burst...

Again, if a correction is not audible, why make it? At the risk of making occasionally some damage? I realised that in most cases is like you said, but there are also cases in which apparently useful material is removed. What is audible or not is also subjective. I don’t have a trained hear and, as I already have a certain age I cannot perceive anymore frequencies above 13-14 kHz, however I am in favour of a conservative approach.
Perhaps an evaluation of the power of the suspected noise vs. the rest of the musical sound should be done, and only those corrections that appear to be beneficial performed.

Quote from: Paul Sanders on 2009-03-01 18:53:00

Couldn't agree more.  Adobe wrecked it.  I am lucky to have a copy.

Cheers and do PM me - Paul.

? I have the last CE and Audition 1.0 -- they seem pretty similar.  how did they wreck CE?

Ok Paul.
I also read again what you and other authors wrote about FFT filtering, and although it appeared (and still appears) magic to me, I must acknowledge that it can introduce some distortions in certain circumstances.
On the other hand, in any case linear filters cannot do serious damage (except for clipping, which, if not negligible, can be avoided with simple rescaling).
Therefore, as I know quite well the kind of noise that I have, which is quite stable in it’s main components, I have just tried a restoration where, besides removing clicks with your program, I applied only linear filtering. I used a series of broad notch filters at 15 Hz and 2x25Hz, plus a narrow notch at 100 Hz. The result is impressive as the remaining noise is practically inaudible (< -74dB), thus hiss filtering can be avoided.
So this is in effect a viable alternative to NR with a noiseprint.

Sonic Foundry and VinylStudio

Quote from: Paul Sanders on 2009-03-03 12:52:33

Sonic Foundry and VinylStudio

I thought it was, at least, competitive - which given the price difference, is stunning.


I haven't had time to play any more since our last discussion. When I do, I'll try out some audio samples that have defeated me (well, SF) in the past.

Cheers,
David.

Quote from: krabapple on 2009-03-02 16:36:09

? I have the last CE and Audition 1.0 -- they seem pretty similar.  how did they wreck CE?

By removing it from the market and then charging 5 times as much for the replacement.

As Clilve said, there is no no good, affordable audio editor (AFAIK).

I've not actually tried it, but I have heard on the grapevine that the low cost version of Sound Forge ("Audio Studio") includes the Noise Reduction 2.0 plugin bundled in.

just to add my 0.02p

Audio Studio doesn't include the NR 2.0 goodies but it does include a vinyl cleanup offering. Still Audio Studio may be a good replacement for CE2000 - the NR features were always an add-on for that anyway at an additionl £40 or so if memory serves

Mine belt-drive TT has had one for years, a low level (mainly audible on headphones during quiet parts) hum only when the stylus is in contact with a record (and only when the motor is running).  I'm guessing that means some sort of mechanical resonance, but no one was able to fix it (this was decades ago when I last took it in for fixing to some hack who was running a local audio salon; he claimed it was fixed but of course when I got it home, it wasn't).

Gawd. People still have hum issues with turntables?

GlaMas:

1) As I said Wave Repair gives the best results in noise reduction (NR). You have an adequate control of the process, in that you can set accuracy of the FFT, you can select precisely a noiseprint, also by examining its frequency characteristics, and you can enable amplitude dependent filtering, something not found elsewhere. With NR done this way (at 1.5 sensitivity) I didn’t experience any undesirable effect.
This is also more than adequate for hum and rumble reduction. In fact, although I read recommendations (from the Authors of WR and VS as well) to use specific hum and rumble filters for this purpose, I found that the response of the notch and high pass filters available is not step enough (at least in WR), so these filters take out more of the bass notes than general NR does.

Paul Sanders:

2) With regard to hiss, hum and rumble filtering, the hiss filter (which is a broadband noise filter) is not a good choice for removing hum as you will lose bass notes if you fail to turn the hum filter on. This is because the hum filter is a narrowband 50 (or 60) Hz notch filter whereas the hiss filter works over a broader frequency range. By enabling the hum filter, you remove the hum from the 'noise print' used by the hiss filter, which in turn reduces the impact on bass notes in general.

3) What I noticed during testing that a quietly plucked bass string almost disappeared when the hiss filter was used to filter out hum, but was restored when the hum filter was turned on. Another curiosity which might pique your interest is that rumble can cause the hiss filter to generate faint clicks (I think because the low frequency component introduces small DC offets which the overlapped FFT's cannot entirely eliminate). Turning on the rumble filter eliminates this and we do it by default when you enable the hiss filter.

GlaMas:

4) Hmm... I didn’t notice anything like that. Can This depend on the accuracy used in the “hiss” filter?
WR offers various degree of accuracy, as well as amplitude dependent filtering, which should minimise undesirable side-effects like this. Isn’t it so?

Paul Sanders:

5) Possibly, but the point is really moot as you have a better option. I don't know what WR's accuracy parameter means, but it probably means a longer FFT length. What I do know is that hiss filtering is inherently subtractive and you don't want anything in your noiseprint (which is what the subtraction is based on) that you can get rid of in a more selective way. Our hum filter (and no doubt WR's) is a steep-sided notch filter which takes advantage of the fact that we know exactly what the mains frequency is and can therefore target it precisely.

6) It may also be the case that overly long FFT sizes are less effective, I'm not sure.

7) Our hiss filter is inherently amplitude (or, rather, SNR) dependent. Again, on the basis of listening tests, this stood out as the right thing to do, although what we do may do it in the same way that WR does. Loud passages of music are certainly much less affected by the subtraction process than quiet passages. Soft cymbals are a particular challenge, for which we have a 'quality' slider which lets you determine, in effect, how quickly the hiss filter reacts to changes in SNR. The bass note I mentioned was quite soft, and given that there were bass components in the noiseprint, the hiss filter thought it was fair game.

One thing that really upsets hiss filters is clicks, so take these out first (VinylStudio always does things in that order) and make sure there are none in the noiseprint. ...
I would *not* apply the hiss filter to a recording that doesn't need it because the FFT .. inverse FFT process (subtraction or no) is inherently imperfect with a finite length FFT.

GlaMas:

9) Ok Paul, thanks a lot.
You almost convinced me not to use the hiss filter and just apply the rumble and hum filters.
I will make some more comparisons.

CliveB

10) I don't know what WR's accuracy parameter means, but it probably means a longer FFT length.
That's right. Wave Repair's accuracy settings mean an FFT length of: Low=2048, Medium=4096, High=8192, V.High=16384.

Paul Sanders:

11) We do overlap our FFT's (of course!) and the clicks introduced in this way are very faint, but they are there. I have some recordings with very high amplitude rumble on the leadin (I think it might be an MP3 artefact, actually) which is probably why I picked this problem up. The rumble filter is applied *before* the hiss filter, and my private theory is that the large low-frequency component increases the discontinuities at the edges of the FFT windows to the extent that the overlap cannot mask them. With the rumble 'pre-filter' turned on, this low frequency component is removed before the hiss filter processes the audio, so the problem goes away.

12) Yes, I can see that using notch filers could be more effective at removing rumble if it has well defined frequencies. Trouble is, it often doesn't. In the general case, I think a high-pass filter is fine. We cascade three, rather than two, rolling off at 32 Hz, 2 octave bandwidth (make it too sharp and it just rings). We found the third filter made a noticeable difference to the sharpness of the response.

13) I just checked our hum filter: notches at 50 (or 60, etc), 100, 200 and 300Hz, bandwidth 0.05 octave in all cases.

14) Hum filter: yes, we could (and perhaps will) offer more control over the notches, but my own personal experience, based on listening to the results, is that the ones we use do not affect the music in any way.

GlaMas:

15) Ok Paul. I also read again what you and other authors wrote about FFT filtering, and although it appeared (and still appears) magic to me, I must acknowledge that it can introduce some distortions in certain circumstances.
On the other hand, in any case linear filters cannot do serious damage (except for clipping, which, if not negligible, can be avoided with simple rescaling).
Therefore, as I know quite well the kind of noise that I have, which is quite stable in it’s main components, I have just tried a restoration where, besides removing clicks with your program, I applied only linear filtering. I used a series of broad notch filters at 15 Hz and 2x25Hz, plus a narrow notch at 100 Hz. The result is impressive as the remaining noise is practically inaudible (< -74dB), thus hiss filtering can be avoided.
So this is in effect a viable alternative to NR with a noiseprint.

Paul Sanders:

16) I think you have made the right decision; it's horses for courses, as we say over here. I don't think clipping is an issue in practise. The hum and rumble filters don't need much headroom, unless you make the high pass filter too steep - it's better to cascade a few shallower ones. I'm not sure if that applies to notch filters though, I've never checked. Anyway, I still favour the HPF for rumble as it does the job just fine.

We are having some beautiful sunny weather here and I hope to get out in it.  You too Glauco, if you don't mind me saying.  Please don't overdo it (it takes one to know one!)

Best - Paul.