Best way to tackle this amount of noise?

[Aya_Rei]

Hello, while going through VHS tapes from a late family member, I stumbled across a tape that contained a recording of episode 5.92 of the Tonight Show with Jay Leno recorded on May 20th 1997. While all the other VHS tapes that were TV recordings were not anything special to me personally, and therefore I didn't bother digitizing them at all. This one is a bit more important to me as said family member was interviewed on a segment during the episode.

Small example clips have been provided, this section of the tape (50 or so minutes, as it then cuts to a recording from Lifetime recorded in SP mode for around 20 minutes before cutting back to an NBC recording) are seemingly recorded in either EP or LP mode, I suspect LP as the neither JVC SR-VS30u or SR-MV45u didn't show a recording speed on the small LCD screen.

If it'll help, the VHS tape itself was a Scotch High Standard T-120, judging by the low audio quality it was also probably recorded from a mono only VCR. Any help with restoration pointers will be appreciated, thanks c:

[timtape]

I'm guessing it's an LP linear recording as it sounds too good for EP.

I've heard LP captures a lot worse! The brief speech in clip 1 is already perfectly intelligible.

It can never be made hifi but a cleaner capture would maybe reduce the hiss in the speech band. The audio head azimuth tracking could perhaps be tightened up a little. The only way to be sure is to capture it optimally and compare the two captures.

As for post restoration I'd probably low pass filter at a corner frequency of about 7 kHz. Maybe also some "presence boost" of frequencies between say 2 kHz and 6 kHz. Others may do it differently. But as with the picture, capturing the audio as cleanly as possible off tape is fundamental to the best final result.

[Selur]

Just for fun, I threw DGDenoise, DeHalo_alpha, YAHR, CodeFormer (<- machine learning, so slow)
script: https://pastebin.com/2cEPGW95
at the clips and it produces something 'watchable', but I would advise recapturing if possible, to archive a better base quality,..

[Aya_Rei]

Thank you so much for you responses, especially yours Selur. While I'd love to recapture this, I already tested the tape in both JVC S-VHS VCRs that I got from Lordsmurf, line TBC enabled and a frame TBC in the chain, with the results from both VCRs looking basically the same, so I don't went to shell out more cash on buying a Panasonic AG 1980 and seeing how the tape responds to that specific VCR.

I've ran countless other tapes that weren't as bad as the Jay Leno clip, so I really much suspect the noise is unique to that tape and other ones like it, I did view a small clip from another Scotch branded tape of the Jay Leno show and the amount of noise was basically the same.

Just as an example really, here's a clip from the same Scotch tape, this time clearly recorded in SP mode as the VCR was able to detect the recording mode. I think you all can agree that while the clip has it's own issues, it looks and sounds better than the Jay Leno recording.

[aramkolt]

If you're curious how it does in an AG1980, after you've captured your tape to the best of your ability (ideally one of the tapes that isn't important to you), you could send me one of the tapes (if you're within the USA) and I can digitize samples of the same area and you can see if the results seem significantly different to what you're currently using.

Odds are that our capture cards or TBCs won't be the same, but I do have several SR-MV series VCRs that I've refurbished, so I can do that comparison versus yours as well.

PM me if interested.

[billuke]

Not to hijack this thread but...
This question is for Selur. The script you included, is that a Hybrid script? If so, how do I load it?
Am impressed with your results.
If I'm breaking any rules here LS... Please delete and
I will post on hybrid forum.
Thanks , Bill

[Aya_Rei]

Decided to recapture part of the tape with the SR-MV45 and compared it to the SR-V30U capture I did first, it's a wee bit blurrier and the linear audio quality is a bit lower. But overall the noise is still present, and the differences seem rather minor, if not in nitpicking territory
https://imgsli.com/Mjg0MDIw

[Selur]

@billuke: Yes, it's a script created by Hybrid. It's not meant to be loaded into Hybrid (one could in theory, if one adjusts the paths and makes sure all the necessary stuff is installed). It's created with Hybrid (+ torch-addon), then Codeformer + DeHalo_alpha + YAHR were enabled and their settings adjusted.

@Ava_Rei: maybe cleaning the headers of the vcr will help,..
others (hopefully more knowledgeable) can probably wage in on whether this should help (haven't touched a vcr for ~20 years)

[Aya_Rei]

Perhaps, but I really don't want to do that as I fear that I'd end up messing up and ruining my VCR(s) in some way, never touched the internals, so all of that is completely new to me. I'm only comfortable-ish enough when it comes to trying to preform restoration work to a video file, and see if it's getting captured properly enough to begin with.

And about the script, I see.. I thought about recreating it myself but I guess I'd need to get a beefy workstation laptop to make use of DGDenoise and torch-addon, assuming Codeformer is included in that? Was able to replicate everything else from the script myself.

[Selur]

DGDenoise should be rather fast on most systems, but requires an NVIDIA card, but CodeFormer is slow and requires the torch-addon, which would probably need a fast NIVDIA card to be usable.
I did play around a bit more (shifting chroma, fixing edges,...):
script: https://pastebin.com/KKVsGRLC (could probably still be tweaked more; also maybe adding some grain at the end,..)

Side note: main thing that bugged me were the horizontal lines, but using MiniDeen helped there.

Cu Selur

[aramkolt]

Selur is the master at work right there! Alas, creating a script like that from scratch is a little intimidating, but it's amazing to see that these tools do exist.

I wonder if there will ever be an AI that'll handle the script writing? Seems like what you'd want is for AI to suggest like 4 different starter scripts of fixing something, then choosing what looks visually the best from a short preview/snapshot of each, then have it make another 4 suggestions as a sort of fork from the last user selection until you arrive at a final script maybe? Of course it accepting text queries to send it in the right direction like denoise/dehalo/add grain with short samples of each level shown would necessary too. Though I would think that AI should be pretty good at identifying halos/ringing/noise is and it would probably get the scripts relatively close on the first couple of selections?

[Selur]

Haven't seen any machine learning or large language model that would be capable of writing proper scripts for a given source.
I doubt there is anything out there that really can do this.
Don't even know of something that is reliably able to identify blended frames or even interlacing, so identifying and differentiating between details and noise seems to be far away.

[billuke]

Thanks Selur.
I do have more hybrid questions, so I'll join / Post more over on that forum.

[timtape]

Quote:

Originally Posted by Selur

Haven't seen any machine learning or large language model that would be capable of writing proper scripts for a given source.
I doubt there is anything out there that really can do this.
Don't even know of something that is reliably able to identify blended frames or even interlacing, so identifying and differentiating between details and noise seems to be far away.

Yes I agree but maybe that's not what some people want to hear!
So in assessing their restoration work operators can thrill at the reduction in noise they've achieved, but at the same time can be blind to the loss of wanted picture information.

In these before and after Jay Leno clip stills, in one the woman's gold necklace appears to be patterned, and in the other plain. So which best represents the real necklace?

[Selur]

in the necklace those look like artifacts, not details to me, but without knowing what the necklace really looked like its naturally just a guess