I really do wish I had the restoration tools back when we started into business 30 years ago. All I had was basically equalisers and a few rudimentary plugins on my DAW.

I had one project that today would have been relatively easy today but was a real problem "back when". This was before the days of cell phones. A man called his wife from England on a landline. She recorded the conversation on her "pocket cassette deck". She was terrified when he told her that he was going to "kill her" when he came home. The conversation was garbled to say the least and her holding the phone and the recorder did not improve the situation. She brought the cassette to us to "clean up the voice" so she could go to the police. I spent two days on it and finally cleared it up enough to make out the words. She took it to the police and the police and FBI were there when the plane landed and they arrested the husband.

Today what I did would probably take me two hours or less and it would be a lot more "cleaned up".

I think it is a marvel what has been "invented" or "created" in the past few years and only wish we had those tools back when...FWIW

I cannot wait to see what is right around the corner when it comes to restoration. The possibility of AI helping is something I eagerly await.

Why wait? I'm already using AI based tools fairly extensively although maybe what I do is somewhat different to many.

Can you tell me more about what AI tools you are currentlhy using. Here it is RX and Spectral Layers and Ozone. Thanks in advance.

As you've probably seen in my other posts, I've been particularly interested in source separation over the last few years and I have just recently tried the beta version of the Ulitimate Vocal Remover which works with a wider variety of algorithms compared to the standard release version.

There is now a pretty decent reverb removal algorithm and also a drum separation algorithm that splits a drum track into bass drum, snare, toms, hi hat, crash cymbals and ride cymbals. I wouldn't be surprised to learn that there are also a few other useful new algorithms that I've not discovered yet.

So now, when I'm presented with a less than perfect mix with no option to remix (as happens fairly often here), I can actually change quite a few things. There's one project coming out in the next few week which used these techniques fairly extensively.

Thanks much!!!

I have tried some of the AI that is suppose to allow one to remix material and so far have not found one that doesn't do some 'WONKIE" things to the instruments and voice. I keep getting request for removing the voice from recordings for KARAOKE use or someone wants to add their voice over the instrumentals. So far no real "luck" in doing so. I will keep on trying...

The best tool for separating vocals that I've found is

https://colab.research.google.com/gi...23-Colab.ipynb

which is probably easier to use than it looks. I was using it to extract a vocal track and it was the most natural sounding extracted vocal that I've yet heard.

Most of these tools will sound perfectly natural if you aren't trying to use tracks on their own. I find that you can often raise or lower things by as much as 10dB without things sounding unnatural.

However, if your experience is limited to RX's Music Rebalance then that's nothing like the current state of the art.

THANKS!!! I will download it and try it. Thanks for the heads up.

I dont necessarily share the same sentiment in the title. Audio transfers done 20 or so years ago should normally be fine if done well at the time. The top AD converters have improved over the decades but for anything but the most demanding tasks they were already pretty damned good back then, and the limitations were almost always in the source material, rather than the converter. So long as the original unprocessed capture has been preserved, which is recommended practice, we can always go back to it and give it the benefit of say an improved denoiser tool or whichever restoration tool has undergone significant improvement since the original capture was made.

Not really talking about A to D converters nor playback equipment since neither of these has anything to do with AI.

More interested in new software AI applications either current or for the future. Thanks for the reply.

Yes but you missed my point. Today's and tomorrow's tools can be used on yesterday's good transfers.

I've found a number of online examples which have something in common : the admittedly lower level of a section or phrase of voice audio is still perfectly intelligible before processing. Here's a recent example:

https://youtu.be/-Tsskh-nhW0?si=fxvzQzOzJjbUOWJN


Here's a professional track by a major artist from 1970 where the vocal starts out with every word quite audible but later, at 4:06, the vocal becomes almost buried underneath the building instrumental crescendo. This appears to be a good candidate for a remix to boost the lead vocal. Even a clean 3db boost in the lead vocal would help.

https://youtu.be/9kxY7Wyzbxs?si=IkL9knPlCEMcgc4E

I got your point.

Right now we are engaged in re-transferring and cataloging a lot of reel to reel tapes done in the 70' and 80's. Some of the original transfers to digital were done OK some not so well. We are, in many cases, having to go back and retransfer the tapes.

Not sure what the original transfers were done on (maybe an Ampex AG440B or an MCI JH-110C or a TASCAM BR20) but for our transfers we are using my Otari MTR-10 II and our new TASCAM 3000SD unit. So far the transfers are coming out GREAT!

A lot of the tapes were done on back coated tapes so we are having to bake them before transfer.

As you said in your post the transfers are the most important step in the process and that is why we are rdoing the ones that were not done well.

I will be using software to do noise reduction and re-eqing the tapes if need.

FWIW

Out of interest, what seem to be the main problems with the older transfers which justified the retransfers?

Very low levels, hum, buzz, distortion and lack of high frequencies.

I played back a few of these tapes and they show none of those problems. Someone was just not doing a good job.

I find this a lot. Someone never does a setup/alignment, sticks on the tape, never checks their levels, plays back the tape and says is it "transferred" without checking the results. Why people do this remains a mystery.

We just got though doing a lot of work on 232 separate recordings and at least as 3rd of them were poorly done transfers to digital. We did not have access to the original materials so we had to work with what was given to us.

FWIW

If you listen carefully to that, you'll notice that it isn't an instrumental crescendo but a crescendo of backing vocals. Consequently the current AI tools wouldn't be able to bring out the lead vocals separately from the backing vocals but feel free to give it a try with UVR (it is free) - you may well learn something if you try it for yourself.

Well it seems a combination of the orchestral backing and the backing vocals with the BV's coming in and out at points after 4:06.

4 years ago I tried a demo of RX Music Rebalance and found the same problem you mention that apparently still afflicts the latest and best tools. There the 50 voice choir I had recorded live in June 2011 completely threw the tool. It couldnt recognize the choir as such. It was with that experience behind me of that tool's limitations 4 years ago that I thought of the Campbell track which apart from the BV's has a complex instrumental backing. I chose the sample precisely because I suspected it would be difficult if not impossible to correct. Thanks for confirming that still seems the case for the current crop of source separation tools.

If your only experience of these tools is an old version of RX then you will find that they have progressed a long way since then. Rather than your normal behaviour of expecting others to do the work, why not download UVR and gain some real experience of the newer tools?

I'm more interested in significant breakthrough points which would normally be made known on forums such as this because they are significant breakthroughs.
3 years ago Peter Jackson, said of their working on the Get Back movie, "We've made some huge breakthroughs in audio... we taught the computer what a voice sounds like. In fact we taught the computer what John sounds like and what Paul sounds like". Not sure what Jackson was actually saying there starting at 1:00 and how it relates to the current state of the art in audio source separation, especially multiple voices.

https://youtu.be/qQFPnA8OgLo?si=QlIKlnDp82R-b7SU

He's probably referring to the ability to train their own models. The problem that many software developers have is that they only have access to a limited set of sample tracks to train this software on.

Somewhere like Abbey Road will have access to a wider range of multitracks so they can refine their models. They can also train the models for specific purposes - in this case working out the difference between John and Paul's voices.

It is possible for anyone to train these models as most of them are open source and available on places like Github. It is slightly frustrating for me as I now have access to quite a few multitracks but I'm not sufficiently knowledgeable in Python and Pytorch to set up a training system.

I understand that the Abbey Road people also use conventional manual spectral editing to separate out voices and instruments. This would be very slow and painstaking work but I guess it can be tweaked to produce better results.

Thanks for the details. I'm basically a hardware guy, definitely not a coder, and I'd prefer to use fully fledged tools such as RX or its competitors as I suspect would most others. I'd also like occasionally to learn what is the state of the art or even Proof of Concept with audio source separation via audio example files. If you or anyone else can link to some audio examples especially (after having trained the software) separation of multiple voices, I'd appreciate it.