I have been trying to digitize my large cassette tape collection and the tape hiss noise removal step is done using Audactiy. About 85% of samples, the Audacity NR tool works great, however there are some samples where Audacity NR actually creates very particular noise.
I have tried many parameters combination in the noise removal tool, but I still can not get rid of it.
Here is a sample (see attached). From the very beginning, the right channel will have a very particular noise (sounds like sands dropping on floor), while the left channel is almost perfect. My theory is that the noise profiling is not very good, and those ‘sanding’ noise are just left over tape hiss. To avoid it, I will have to set Noise Reduction (dB) to a much small number (like < 6), but then tape hiss will be mostly preserved; I also find that a large ‘Attack/decay time (sec)’ sometimes may help, but only to spread the ‘sanding’ noise, while ‘Frequency smooth’ does not seem to have any effect.
Did I do anything wrong? What is the best way to find a optimal NR parameters? Any comment?
If anyone wants to experiment, here are the original wav files:
I’m surprised you got any hiss removal to work. Noise Reduction works by automatically generating rejection filters based on whatever you used for the Profile. That’s why the Profile step is supremely important. Since hiss is a sound made up of all musical tones, Noise Reduction tries to remove the whole show.
Frequency Smoothing tries to tell the removal process when to stop. In modest circumstances, you can get Noise Removal to stop working during spoken words thus preserving the quality of voices while still removing noise between words. The smaller the number, the less the effect. If your show is a blizzard of hiss, this will just give you hissy words over a clean background – or other settings will give honky, science fiction words.
I need to listen later, but chances are if you couldn’t do it, we won’t be able to, either.
Well, the sample is only one of the 14 tracks from a cassette tape. Rest of the tracks are mostly very very good after NR operation. Even for the problematic track I posted here, the left channel is also very good after NR, and only the right channel is bad.
I have used several cassette decks (all are Sony TC ES decks) as analog source, but this particular right channel ‘sanding’ sound are all there after processing. The major difference of this track, comparing with rest of the tracks, is the music starts in a relative lower dB. That is why I suspect there are issues with the noise profiling steps, maybe the noise collected is not appropriate.
What happens if you click in the Track Name, choose “Split Stereo Track”, take the noise profile from the left channel then noise remove both channels?
Or if you profile the left channel, noise remove the left channel, then profile the right channel and noise remove the right channel?
It turns out the noise profile using only left channel (of course, I have to play the split stereo trick as you have suggested) improves the result a lot. It is not perfect, but the artifact is reduced a lot, though still audible at some spot.
So it looks like the noise profile is indeed the issue.
Is there a way to analyze the spectrum, and recreate the noise for noise profile?
One trick I use to make sure I’ve just got “noise” is to amplify it up a long way with the Amplify effect and then listening to it. Sometimes I will find a very quiet fade out/in included as part of the “noise” - which means that some musical content would be removed if I used that “noise” sample. Then Undo the amplify before using it as the “noise” sample for the first pass of the Nose Removal effect - or try to find a better pure “noise” sample elsewhere in the recording.
I also find it hepful to temporarily label the “noise” sample while working on it.
Another trick is to test the noise removal on a shortish section of the music (or whatever) to make sure that it sounds ok after noise reduction before committing to noise removal on a long recording (removing too much signal can make the recording sound dull and lifeless).
I think Gale’s suggestion, in your case, of doing separate noise reduction on each channel, each with its own noise sample, is a very good one.
I also found I got better results by slightly softening the noise reduction from the default settings (following a suggestion from Steve).
This is a note I added to the Noise Removal page for the alpha 2.0.6 manual:
Alternative settings
For heavier noise, the default settings above are recommended, increasing the “Sensitivity” slider if necessary.
For light noise (for example FM hiss or web-stream carrier noise) the default settings in Noise Removal may be a little too aggressive, possibly introducing tinkly artifacts. Noise Reduction (dB): Try setting this to about -12. It will still have an appreciable effect on low level noise, but will reduce the risk of tinkly artifacts. Frequency Smoothing (Hz): Try increasing this a little, typically to around 500 Hz. Sensitivity control (dB): Reducing this can further help to make the effect less aggressive, but may require a lot of experimentation for only slightly better results. For a quick result, this slider is usually best left at the default 0.0 dB.
For critical work it is best to experiment for optimum settings; see this page in the Wiki for further advice on using the settings for this effect.