Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler" (Read 7691 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #50
At some point I may consider getting a new GPU and learn some Python then run the code locally on my PC, and do some listening tests myself.
The CPU version of this do already exist but expect worse performance (perhaps the only option if you don't have enough money to buy a NVIDIA GPU) than GPU-accelerated version

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #51
No. My opinion is in the quoted text above. My statement may contain grammatical mistake but I did not say "null test is the way to go", just to be clear.
More cryptic content than grammatical mistakes in your text, then.

So you're for ABX (= listener's personal perception) that AFAIK is mainly used for compression quality evaluation.

At some point I may consider getting a new GPU and learn some Python then run the code locally on my PC, and do some listening tests myself.
Unneded: if a Jupiter notebook comes out (hopefully soon), everyone can perform own tests.

But - to me - the focus is on neural network TRAINING.
Hybrid Multimedia Production Suite will be a platform-indipendent open source suite for advanced audio/video contents production.
Official git: https://forart.it/HyMPS/

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #52
So you're for ABX (= listener's personal perception) that AFAIK is mainly used for compression quality evaluation.
ABX is used for perceptual quality evaluation.

If the upscaler is supposed to make lossy audio sound more like the original, that's perceptual quality. You need blind ABX testing to measure the perceptual quality.

If the upscaler isn't supposed to make lossy audio sound more like the original, what's the point?

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #53
ABX is used for perceptual quality evaluation.
That's a stretch. It is used to measure whether two signals differ perceptually. Try two bit-identical copies of an original signal - they are equally good. Try two independent white noise signals - they are equally bad. Try two lossy versions of an original - they may or may not be "equally bad" even if they are easily ABXable from the original and from each other.

If the upscaler isn't supposed to make lossy audio sound more like the original, what's the point?
Make artifacts less annoying
Sure there might be a "reasonable yardstick" that says this makes for "more like the original" - but even if easily ABXable, it might still improve.

Already more than thirty years ago, Photoshop could processing scanned images to remove what likely was distortions in scanning (like, despeckling raster images). The software itself could not know that it wasn't intended to be, but the filter would improve if used wisely - improve, meaning the picture wouldn't look so obviously distorted (that is "less annoying"), at least to a viewer who would not have the original to ABX against, and neither would the person doing the image processing.
Artificial intelligence can make a pretty good shot at "used wisely".

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #54
That's a stretch.
You're right; I think I was mixing it up with ABC/HR, which is also a method for perceptual quality evaluation but geared more towards "which one sounds better" instead of "is there an audible difference".

Make artifacts less annoying?
And in that case, you still need perceptual evaluation - like ABX or ABC/HR - to decide whether you've succeeded.

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #55
That's a stretch.
You're right; I think I was mixing it up with ABC/HR, which is also a method for perceptual quality evaluation but geared more towards "which one sounds better" instead of "is there an audible difference".

Make artifacts less annoying?
And in that case, you still need perceptual evaluation - like ABX or ABC/HR - to decide whether you've succeeded.

Agree.
But do it yourself. Don't talk about how difficult is to use these novedous AI tools...

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #56
No. My opinion is in the quoted text above. My statement may contain grammatical mistake but I did not say "null test is the way to go", just to be clear.
More cryptic content than grammatical mistakes in your text, then.

So you're for ABX (= listener's personal perception) that AFAIK is mainly used for compression quality evaluation.

At some point I may consider getting a new GPU and learn some Python then run the code locally on my PC, and do some listening tests myself.
There should be no cryptic content that ABX is a type of listening test, but listening tests are not necessarily ABX.

Not specific to Fat Llama, but for example, there should be no doubt that 32kbps MP3 should be nontransparent when ABXed with the original, but I am not in hope that the AI-restored 32kbps MP3 can be transparent when ABXed with the original file.

Let's say someone invited some professional musicians (solo, small band) to a recording studio with controlled recording environment (equipment, acoustic treatment, miking etc) then ask them to perform a piece of music 10 times. Even if all the 10 performances are flawless, it is still possible that individual performances can be ABXed against each other, and putting these recordings into Deltawave won't give you any meaningful result. In case of any doubt it is possible to ask the author of Deltawave, I also asked some questions like this:
https://www.audiosciencereview.com/forum/index.php?threads/beta-test-deltawave-null-comparison-software.6633/page-18#post-295582

Let's say, if 2 or 3 of these recorded files are being put into the 32kbps MP3 -> AI restoration signal chain, can people identify the differences against the other 7 or 8 unadulterated versions?

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #57
If the upscaler is supposed to make lossy audio sound more like the original, that's perceptual quality. You need blind ABX testing to measure the perceptual quality.

If the upscaler isn't supposed to make lossy audio sound more like the original, what's the point?
As sayd the focus should be on neural network training: you can feedback (read improve) it both with personal perceptions and "objective" ones.

I do believe that second ones could be more effective.
Hybrid Multimedia Production Suite will be a platform-indipendent open source suite for advanced audio/video contents production.
Official git: https://forart.it/HyMPS/

 

Re: [OPEN SOURCE] Fat Llama: a (wannabe) lossy to lossless "upscaler"

Reply #58
Let's say, if 2 or 3 of these recorded files are being put into the 32kbps MP3 -> AI restoration signal chain, can people identify the differences against the other 7 or 8 unadulterated versions?
Maybe I'm implying this 3ad purpose cryptically, then: I've opened this discussion here at HA (where I know there are many experts I respect) not mainly to evaluate the quality of the results that fat llama actually produces but, since it's open source, to understand/suggest @bkrd47 how can be improved...
Hybrid Multimedia Production Suite will be a platform-indipendent open source suite for advanced audio/video contents production.
Official git: https://forart.it/HyMPS/