Thank you so much for your efforts, lucianodato! I’ve just tried the latest git master, and noticed some minor things to improve:
Residual whitening generates repetitive noise patterns, which is no good as I’ve found some situations where those patterns can be easily heard.
There’s no frequency smoothing parameter (like in Audacity), which is very useful to clean some non-musical sources (e.g. nature sounds, ambience, etc.), especially those with useful background noises (e.g. distant traffic, rain rattle, etc.).
No way to set FFT window length. I’m not sure whether there should be such a parameter (as well as “overlap”) but there can be some situations when shorter FFT windows produce better results, as those are less aggressive, especially to sounds short attack times. In fact, it’s a good field for experiments, I mean to devise some way of dynamic adjustment of the FFT window length according to a differential threshold (i.e. attack length). I know it’s probably very complicated to implement, but could produce some fine results.
The plugin does not work correctly in Audacity, while it would be really awesome if it could.
Thanks @Efenstor for your comments. Here are my answers:
1- Residual whitening is badly implemented and I know that. It’s in the list of issues in github. Spectral whitening should preserve the envelope of the residual signal and reduce only peaks and nulls that are far from it. What is doing now is wrong and only works for brown or pink alike noises. I’m already working on fixing this.
2- Frequency smoothing often sounds bad because it tends to reduce lower frequencies in order to avoid musical noise but I didn’t thought of your use case. With the new masking model it seems to do much better in all the cases I was testing it on. Maybe with some explanation of what parameters do you’ll find that is effective for your situation (Will update DEV branch readme soon).
3- You mean to adaptively resize STFT blocks based on transience or something like that? Ala izotope? This indeed would be a great improvement but would increase cpu usage a lot since It would involve processing various FFT sizes at the same time and then combine them. It’s doable but maybe not for realtime usage though.
4- Already on github issues. No clue how to make it work. Audacity seems to destroy every variable in memory once the playhead stops. Again not really sure of the ins and outs of this. Maybe I could chime in in audacity forums and ask.
If you have an example where previous versions of noise repellent were working better than current one don’t hesitate to send me those. It would be great help.
Hi Lexridge! I’ve implemented a perceptual model to even get more and cleaner reduction. It’s sounds amazingly well. That is already up in the development branch in my github. I think that is reaching pro level now.
@ luciano you are fast! I will test soon! your plugin already helped me a lot on some productions. at the moment “unfortunately” i have no noisy dialogue I could use it on but will test soon!
Better yet, I have changed the dest directory in the Makefile from PREFIX ?= /usr/local to PREFIX ?= /usr. Should I also change the LV2DIR ?= $(PREFIX)/lib/lv2 to LV2DIR ?= $(PREFIX)/lib64/lv2?
Just built the latest from git. I will give it a try later in the week. I have noticed that you default the install to /usr/local/lib/lv2 and I always move it to /usr/lib/lv2. I am curious, is this doing a 64 bit build on my 64 bit system? Perhaps I should be moving it to /usr/lib64/lv2 instead?
This plugin is excellent right now, and an absolute necessity for anyone who ever has had a little, or even a lot of unwanted noise in a recording - that describes everyone! I successfully downloaded Noise-Repellent from your github, and had it made and running on my AVLinux Ardour NUC in less than 5 minutes. Well done, and thank you!!!
@paul I would love that. How would it work for new versions of it? Will you update every release of Ardour? If not I would prefer to finish current version that implements an overall better algorithm. Would you guys take it as is? Do you need me to document the code better using doxygen? Might need to rethink some stuff to follow Ardour coding style better maybe.
Yea time is always key, been trying to get my consulting business off the ground and currently am gearing up for a couple weeks of tech of a production of Legally Blonde, always seems to be something, but seriously, thank you for working on this. Restoration has been a serious weak point for audio processing on Linux, I really do look forward to hearing some of the improvements you have made!
Me too. No time presently to try it, but I will indeed. I have been very involved with a system install from both Harmonic and Crispin automation at work which has left little time for fun and lots of unpaid overtime, which sucks! Twelve to fourteen hour days is all work and no play. Two more weeks of this. Grrr. Just a word of advice…if your job requires you work with Harmonic Corp, beware of bullshit, because they will deliver tons of it. I can buy a personal $5 dollar lifetime MPEG2 license from the MPEG consortium, but Harmonic wants $750 for it, per channel ( and we have 16) . Fucking bullshit! Legal crooks! Sorry, still ranting from my day at work lol.
I will test soon! your plugin already helped me a lot on some productions. at the moment “unfortunately” i have no noisy dialogue I could use it on but will test soon!