I have some demo material from a UK band that was on cassette and transfered to CD. I do not have the cassette available as I received only a CD-R copy. I have been trying to remove tape hiss with Sonic Foundry Noise Reduction Plug-in. For some of the tracks I can get a decent noise print and can remove the hiss pretty readily. I have tried to remove it in a few passes 2-3 or many many passes 20-50 0.1 - 0.2 dB at a time. However the tracks often sound very heavy on the bass after removal. I would love to hear any suggestions that might improve the sound. I also have some other material that despite being more recent has been transferred to CD more poorly and has some clipping as well as a lot of tape hiss and possibly some issues with DC offset. I have started work with this too much yet since I have been focusing on the other but I would love any suggestions for this situation as well. Finally I'll note that I have a friend how has the suite of Waves plugins so if there are tasks it might handle better I can head over there.

I have had good results against tape hiss with these settings in the Sonic Foundry NR plug-in. Results are best with a noise sample of adequate length, which I would say is four to eight seconds. If you are unable to get a long enough selection of noise, copy and paste several different (very similar) segments together in a new file of just that noise, in order to make the noise print. Or copy the same selection several times side by side to create a long enough duration.

When you need to paste together a number of smaller pieces of noise-only audio, adjust the selection boundaries to zero crossings before you Copy. This will avoid transient spikes where the pieces are Pasted together.

Make sure the noise is really only noise by listening to it at extra high volume, which should reveal low level music (bad) if it exists in your selection. Sometimes for hiss, if you cannot get a selection without any hint of music, a high pass filter, cutoff at 500Hz or so, may remove most of the music audio while leaving most of the tape hiss.

Sometimes it is worthwhile to repeat the entire process again, or even twice, after the NR pass.

on the General tap
reduce 10dB
noise bias 0
attack speed 100
release speed 100
FFT size 16,384
Overlap 75

on the Noiseprint tab
mode 3
Fit Size 2,048

you should try waves' new noise reduction plugin z-noise.

Are you sure that, copying the same selection several times in order to make longer the noise selection, is useful?

If the program has more data it can produce a better plot (number of points in Fit size). If there just happened to be a longer section of tape hiss, would it be much different in content than a shorter section? If not, then duplicating that shorter section a number of times would not produce something different than the longer section. The program will produce a better profile from adequate data than from too short a selection. The profile curve will look different and the results will be a bit more "surgical."

I had the same impression. For hiss removal a sample of about 100 ms should be enough. Splicing samples together doesn't seem a good idea and might even create artifacts (low frequency bumps). If you're really concerned about the sub 10 Hz region (which is likely to contain only unwanted noise), a longer sample might be useful, but a simple highpass filter is probably more effective in that case.

Thanks for the suggestions. For the CD-R of the first tape I do hape plenty of just tape noise between tracks for the noise prints. For some of the CD-R's of the other tapes I think there will be a bigger problem with regard to getting a decent length noise print. I also wondered is it worthwhile to switch to 32 bit before doing the noise reduction? How about DC offset? The offset seems to be fine on the first CD-R (no changes after running the DC offset but in the others it seems to shift to a large degree.

That impression (from some little birdie?) is of a fantasy, not reality. It is very easy to observe that the noise print graph comes out very different using a selection of 100ms vs. using several seconds. Duplicating a several seconds long selection to make it still longer will again refine the noise profile, though to a lesser extent. This is objective and easily verified.

The fact that such a profile works better is perhaps more subjective that objective. It does discriminate better between the noise (that was present in the selection used to make the profile) and the signal (that was not present in the selection used to make the profile), but whether or not this is actually superior will, I guess, depend upon your aesthetic and emotional preferences.

CoolEdit's/Audition's noise reduction also works better with a higher precision noise profile but it isn't easy (at least for me) to make a direct comparison to the way to Sonic Foundry functions. Audition allows a much higher number of data points (Snapshots in profile and FFT size) than does Sonic Foundry but those numbers don't seem to function is quite the same way.
(a) Audition will definitely improve with number of points far above the maximum that Sonic Foundry allows.
(b) Audition will simply refuse to make a profile unless the selection is long enough to support the precision parameters. Sonic Foundry will proceed regardless, it just does a poorer job when the selection is too short.

Perhaps it is technically incorrect, but I don't consider subsonics as part of tape hiss. That is high frequency stuff.

The very low frequencies on recordings taken from cassettes are generally at about the same level as frequencies at the low end of the audible spectrum (unless there is some problem such as 60Hz hum). The subsonics level on every vinyl recording from every playback system I have ever been able to analyze increases greatly at very low frequencies.

I consider those frequencies from either medium as unwanted and always use a rumble filter to reduce them before doing any NR. The filter on the Sonic Foundry Click and Crackle Removal tab does a good job in most cases. Just move the three control sliders to the minimum value (far left) and turn on Remove low frequency rumble. There are many way to filter out subsonics in Audition.

If one is only doing NR, or NR and just a little other processing, converting to 32 bit doesn't make an audible difference to me. You will need to add more noise (dither) when you reconvert to 16 bits, but that is unlikely to be heard either.

How are you determining that you have a DC offset? Some people are concerned if the waveform is not symmetrical around the center line, but this is no indicator of DC offset and is very often correct, i.e. what the waveform should look like. Modifying such a file to center it will have very unfavorable consequences, especially on the lower frequencies.

I'm pretty sure DC offset in the recording has no ultimate practical consequences but its presence can effect the results of some transforms used in mixing and mastering. Since my soundcards don't have offset problems I am not familiar with which transforms one has to worry about.

Thanks again for the input. When I first started this noise reduction project a few weeks ago I tried to read up on techniques. Several times I ran across instructions that said to first run the DC offset correction. When I do this with the present material there is no change (ie I don't see the wavform change and there is no asterisk showing that a change has been made to the file). So for now I assume it is not an issue. However I have some material I will be working on soon that seems to change dramatically after running DC offset. I didn't transfer the cassettes to CD-R so I'm uncertain if there were problems with the original soundcard or not though I suspect it was done 6 or 7 years ago so I imagine soundcards have progressed a lot since then.

I will try the tip regarding running remove low frequency rumble in the Click and Crackle Remover. I will say that running at the settings mentioned above if I listen to the residual output I can hear parts of the music clearly so I'll need to go with a much lower value than 10 I believe. I have gone multiple passes (40-50) with just 0.1 or 0.2 at a time and I don't hear anything in the residual output but the end result is quite bass heavy.

I would say that such small increments are pointless.
However, if you are doing 40 to 50 transforms of any kind, then my statement on converting to floating point doesn't hold. I don't know how much unpleasantness the rounding errors add up to in that number of operations, but I suspect I wouldn't like the results. Definitely use floating point.

So you hear some music in the removed portion? So what? Take the final result of 40 or 50 small operations. Subtract from the original to get just the total ‘noise' that was removed. I suspect you will hear just as much music in that. Of course if you only want 5dB, or 3db, or whatever, total noise reduction, use that number instead of 10. Otherwise, you are unlikely to hear any improvement from summing multiple smaller runs.

Extract a good track from a commercial CD labeled DDD; its never came near tape.
Generate some noise of that length, say pink noise, or record some tape noise off a blank cassette. If you generate noise, adjust its level to about the RMS value of the tape noise you are currently trying to remove.
Mix the two together. You now have a good track that sounds like it came from tape.
You also have a large sample of the noise from which to make a good noise profile. Making a profile from the noise you used to create the mix assures you there is no trace of the music in the noise sample; you have an ideal noise profile.
Run NR against the mix with the parameters I suggested. You will hear some trace of the audio in the residue. This is because tape hiss, white noise, pink noise, etc. coves the spectrum fairly well and can't help but match a little of the music. NR removes that same proportion of music/noise no matter where you set the Reduce noise by slider.

If you get significantly more music mixed with the noise in your real project than you got in the test, you did not make the noise profile with a clean enough noise sample.

The question is not whether or not any trace of the music comes out with the noise, the question is whether or not you can hear any degradation of the music in the finished product. I've done quite a few A/B test on projects in process to be certain that I like the results I'm going to get. I know it works well and I know that I usually can't hear any reduction of the music. Sometimes, when there is enough noise, you have to chose between noisy and something a little different than the original would have been, had there never been any noise to remove. It's your choice, but the optimum place is generally somewhere on the gradient between no NR and no audible trace of remaining noise.

In my (limited) experience, any DC offset is so small that correcting it dos not produce a visible change in the waveform. Do you have software that will measure the DC offset, so you can be sure it really exists and you aren't just dealing with some (normal) asymmetry?

Thanks again for the help. I realize there will likely have to be some compromise between the hiss removal and the music removed along with it. I'm not expecting perfect results I just want to try to do the best with the tools I have available but I have very little experience with this particular task. In the past I had mostly down click removal on relatively clean vinyl transfers. I know very little about DC offset I have Sound Forge available. I just know that following some instuctions that I found on the web suggested to always do this first. It makes no change at all to my current project but the next CD-R I will work with shows a noticeable change in the peaks in the waveform after the operation. I will also test Waves noise removal at my friends so I would love to hear any comments on that particular plugin. Again I know it will not be perfect I just would like to get it sounding a little better for an anthology of studio demos I am compiling for this UK band.

Except for the NR plug-in I don't know Sonic Foundry; I use CoolEdit. Certain basic facilities ought to be much the same, however, because they are so useful. Maybe the following considerations are obvious, old hat, but in case some of them might be useful:

CoolEdit has a good Spectral View and a fairly reasonable Frequency Analysis display. They are very useful for augmenting your listening to provide insight to what your manipulations produce. If Sonic Foundry has similar tools, learn to use them often.

It depends a great deal upon both the composition and the intention of the performers, but recordings frequently have some parts that are at a much lower level than the majority. Sometimes this is only at fade-outs, sometimes there are longer low level passages, perhaps a single instrument or single voice giving dramatic counterpoint to the rest. These low level parts are frequently the easiest places to tell whether your transforms are having good or bad results.

There are two particularly useful methods for observing results before committing yourself. For one of them you select a short stretch, a few seconds or so, in the middle of some passage. Adjust the selection to zero boundaries so the transform doesn't produce clicks at the selection edges. Apply the transform(s) --e.g NR-- to the selection.

Now select a larger region surrounding your transform application. Play through. Do you like whatever difference you hear when the changed part comes up? Don't forget to Undo after your experiment. Obviously you can repeat with different settings until you get the best results.

Make the selection as above but then copy it. Paste it into a new file. Now paste it again immediately following the first paste. Apply the transform to one of the copies. Loop play through the entire new file until you are certain how you feel about the original vs the changed audio.

I do many vinyl transfers. If the LP is in good enough condition there may be no higher frequency noise I want to remove. However, if I listen carefully between tracks and in lower level passages, I can always hear the sound of stylus scraping over vinyl. I generally want to attenuate this noise even when everything else if fine.

I made a preset for a lowpass FFT filter that cuts off at 500Hz. Using the Frequency Analysis graph on some filtered audio, I see that this filter is a real ‘brick wall.' Above the cutoff frequency the signal level drops essentially straight down for more than 100dB. Of course such an extreme slope isn't generally necessary for practical functionally but it does illustrate a couple of things.

Apply this filter to a section between tracks, where no music can be heard at any volume level. Now nothing exists above 500Hz in that selection. Almost always, before and after RMS Average measurements of signal strength differ by less than 1dB -- and this is with a rumble filter applied before the initial measurements, before the 500Hz filter. In other words, the vinyl surface noise is heavily concentrated in the lower frequencies. I lowpass filter the noise sample in order to remove only LF noise.

In actual use I adjust the cutoff of this filter by first looking at a Frequency Analysis Graph of the noise sample. The default is 500Hz, but I set the cutoff somewhere between 400Hz and 1000Hz, depending on what I see in the graph. The relevant item here, however, is what happens when NR uses a lowpass filtered noise sample.

CoolEdit's NR transform does exactly what I tell it. It does not, as far as I can determine, touch anything above the cutoff frequency; Sonic Foundry does. There can be enough HF reduction to hear some difference.

With Spectral View or the Frequency Analysis display it is easy enough to see that music frequencies at the high end of the spectrum are somewhat attenuated after applying Sonic Foundry's NR. Most of the time this is not enough to be a problem, so it does not prevent me from using the plugin. Other benefits of the SF NR plugin often make this small HF attenuation a worthwhile trade off.

I don't know, I've never done anything approaching the 40 to 50 iterations of NR you wrote about (and I don't believe that is a useful approach, but compare both ways and find out), but if SF effects higher frequencies when there are no higher frequencies in the noise sample, possibly it effects them to some similar amount regardless of the Reduce noise by setting. If so, this could certainly explain why your work sounds bass heavy after so many NR runs -- you have unintentionally removed a significant part of the higher audio frequencies.

Sometimes the short selections of sensitive passages I described above do not provide an adequate test of the overall results of some treatment; this may well be so if you want to really test your approach of many small NR applications vs a single, or a few, larger ones. Select something longer, usually an entire track of several minutes or so length. Save it to disk as the reference.

Apply different treatments of whatever is in question to the entire track, saving each result with a different defining filename. Now you have something real to compare in order to give you an idea of which treatment is best. You didn't have to spend the time, or disk space, to test the entire album. I've done hundreds of such tests over the past few years. I believe they are a valuable learning tool.

When differences are subtle, WinABX is a good tool to use. It will sometimes tell you that those small distinctions are all in your imagination, that it is impossible to tell the results apart when you don't know which you hear when. This can be rather interesting, and a little disturbing, when at first you are certain the differences are quite real.