I'll spoil it. Noise removal for early, turn of the 20th century acoustic recordings. Generally, the only audio is piano accompaniment and vocals.
Here is an example. The first is without processing. The 2nd is after I separated out the various tracks with UVR, and kept the ones that had content (I kept vocals, other and piano, got rid of drums, guitar and bass). The differences are stark, and I really can't tell the difference in the vocals between the 2 (both aren't the greatest quality, the singer was belting it into a horn, after all).
The problem with this approach is that it's still manual (I'm not sure that it's a given which tracks you would keep, I would imagine it would vary for each recording, particularly if other instruments are involved - I still don't know how they crowded an orchestra around a horn for some recordings).
Addendum. After listening again, I do note that the 2nd recording sounds a bit hollow or echo'ey (sic). Maybe I'll try some different models to see if that improves. I don't think I'm going to do this for almost 2000 songs and counting, though.
.... kept the ones that had content (I kept vocals, other and piano, got rid of drums, guitar and bass). The differences are stark,..
Pretty impressive results, but I'm confused about you saying you removed the drums guitar and bass. - what drums guitar and bass? I didn't hear any in the original mp3
The model I used split the audio into 6 tracks, those you mentioned. Most of the noise was actually found on the drums track, I believe (for whatever reason).
And the process isn't perfect. Some of the audio (the really high notes) I found on the guitar track. The tracks that I didn't include in the final mix didn't have anything on them but noise.
And most of the piano ended up on the vocal track, with some on the piano track. Weird. I think it might be a byproduct of the acoustic recording. The model was probably trained on pristine, CD audio.
Anyway, it seemed to work fairly well.
I'm getting some pretty impressive results with other acoustic recordings as well. I think I might be able to automate this process. UVR will operate in command line mode. Audacity won't (to mix the several tracks without the noise), but the open source app sox will do that quite handily.
I already use both ffmpeg and sox to automate the preprocessing of my audio files. I cut off leading and trailing silence, normalize, and then encode to a predefined cbr mp3 bitrate. Sometimes the normalizing and cutting off silence doesn't quite work as well as I would like, due to noisy recordings. If I pre preprocess each file with UVR, that might solve the problem, and also make these early recordings much more pleasing to listen to.
@artisan-radio I get it now. I thought your were saying you removed those musical instrumentations from the actual audio composition.
@artisan-radio I just put your cleaned version into Suno and told it to "add drums, guitar and bass"..
The other one it spit back out at the same time, not sure which is better, but figure might as well show the other option
Ok one more time, just playing here, this time I put the one with all the noise in it and told Suno to: "Duplicate original (restore), same female voice and piano and add drums, guitar and bass."
Just wanted to see what it would do.
It even transcribed the lyrics!
[Verse 1]
Heilig’ ist die Morgenstunde
Sollst du Wasser auf und nieder
Und die Abendzeit am Springbrunn
Wo die weien Wasser pletschern
Heilig’ ist die Morgenstunde
Und die Abendzeit am Springbrunn
Wo die weien Wasser pletschern
[Verse 2]
Heilig sollst du sein und leise
Leise und leise
Deines Abends Pfad versen
Auf des Abends Pfad versen
Deinen Namen still zu wissen
Deine Heimat, deine Liebe
Und der Glaube
Wo die weien Wasser pletschern
Und mein Traum sind jene Wasser
Pletschern ewig
Und mein Traum sind jene Wasser
Pletschern ewig
@artisan-radio I listened and the second was much improved and not noticable loss in audio on the voice but I did notice since these were both identical tracks that the second one you took out the piano in parts that was there in the original.
I don't like with AI adding artificial instruments as Rich did as if you are going to do this stuff it has to be authentic. Cleaning up the noise and trying to improve audio quality a bit is one thing but adding artificial things to it to change it from the original recording as it was back then is another.
@mark I was just playing around Mark, just wanted to see what it would do. I realize what Artesian is trying to do. I was just messing about, Suno certainly not the way to go. But that said, AI is definitely the way the professional restorations methods have already shifted to, it's the primary engine now (not music emulators like Suno of course, but ai engines specifically trained in restoration.
The mind is a funny thing. If you listen to a noisy track, it can filter out that noise and fill in the gaps, so to speak, of what might be missing underneath.
If you take away the noise, the missing stuff becomes more apparent.
In the case of the tracks I posted, there was absolutely nothing but noise on the separated tracks I left out, except noise. If anything was missing in the remaining tracks, it wasn't there in the first place.
That being said, I've somewhat automated the de-noising process, and I've found that, depending on the audio content, the separation into multiple tracks isn't exactly perfect. In some tracks, depending on the frequencies of the vocals and the noise, a bit of what you want to keep gets buried in the noise track (usually drums), and you can't do anything about it. It's all about the AI training.
You're left with the choice of either keeping the noise, or losing some of the audio. I'd choose the former, for, as I stated earlier, the brain is an amazing filter.
This has been a good exercise, however. On one hand, it shows the power of the computer algorithms known as AI. But on the other, it shows their limitations, and how much further they have to go. In particular, referencing the latter, it demonstrates how dependent they are on the domain they've been trained in, and how they really can't function outside of that domain.
Which is really true of all computer algorithms, no matter how much you hype them.
OK, I've gotten even better results.
First, I separated an audio track out into instrumentals and vocals using the MDX23C-8KFFT-InstVoc_HQ model, which is supposed to really work well.
Then I took the instrumentals, and ran the htdemucs_6s model, which splits that particular track into 6 other tracks - drums, bass, piano, vocals, guitar and other (everything else).
Virtually all the noise ended up in the drums track, so I mixed the remaining 5 tracks together, along with the vocals track from the first run, and the results were amazing. Virtually no noise, and everything there.
I'm now experimenting to see if I can use the ensemble mode (allowing you to run multiple models) to see if that produces the same result. Not entirely sure how this mode works, so we'll see. I'll post the various tracks once I get the ensemble results.
The mind is a funny thing. If you listen to a noisy track, it can filter out that noise and fill in the gaps, so to speak, of what might be missing underneath.
If you take away the noise, the missing stuff becomes more apparent ..... ..
I know exactly what you mean, and sometimes you might feel like your making great progress working on cleaning a file for hours, but later listen back to what you had been so satisfied with at the time, suddenly sounds so bad that you wonder how you initially thought was an improvement - and then have to start all over again, because all the work you did was for naught. Or at least that's how it was with me years ago when I was still trying to clean the QP episodes.
I'm intrigued with your experimentations using UVR, and just from the samples you posted, your making genuine progress in the right direction from what I can tell.