spc2vgm

[niekniek]

Check overhere for my new collaboration with AI; spc2vgm. It's pretty close to the original I would say, but let me know if things sound off.

https://github.com/niekvlessert/spc2vgm/

I included some vgmrips 'style' packs from converted files as attachments.

[Mildanner]

Can you make a release of it?

The VGM keeps closing my VGM Player

[Kaminari]

Nice proof of concept! The result is very convincing, minus the echo. And the files are not too big for sample-based tunes.

You should mention in your repo how much AI (if any) is actually involved in your project. GitHub is swarmed with unreadable junk coding these days.

[niekniek]

Can you make a release of it?

The VGM keeps closing my VGM Player

Working on it. What player do you use? I don't have such trouble. You can try https://niekvlessert.github.io/vgmplay-js-2/ , you can upload new zip files at the bottom left. It also supports rsn for spc, so comparing is easy...

Nice proof of concept! The result is very convincing, minus the echo. And the files are not too big for sample-based tunes.

You should mention in your repo how much AI (if any) is actually involved in your project. GitHub is swarmed with unreadable junk coding these days.

Echo still needs work, It uses the extra channels of OPL4 to make it convincing, that should be the best method, but it has issues. I fixed something for Lufia, but severe regressions, don't try the latest version on github.

You make an interesting point about mentioning AI. I read a good comment a while back; 'a fool with a tool is still a fool'. I'd like to see myself as a valid software engineer, I studied for it anyway, but I'm too impatient to go deep into the details of languages, libraries, frameworks, IDEs, etc. every day, so I ended up in another part of IT. It also explains why my code is messy probably, I lack hours of coding. But these days I'm very happy with AI; if you ask it the right thing and keep the context for it valid, it'll work fine I'd say. I added an md file to the project for AI models and human developers about the code. However I agree the code is not very readeable now for a human, I'll try to fix that later too. But I think the code itself is not that messy, let me know your thoughts about it. Basically the code itself is generated 100% with AI, I just am architect, product owner, system manager and tester...

[misha1350]

This is BIG. Twinbee Rainbow Bell Adventure sounds pretty authentic to the real deal, minus a few iffy ear-splitting parts scattered here and there (I don't remember them being present in the official soundtrack release). It also hasn't closed on me once on VGMPlay 0.52. I'd prefer a full gamerip, especially since the quality actually looks to be better than what the existing gamerip on Khinsider has.

You make an interesting point about mentioning AI. I read a good comment a while back; 'a fool with a tool is still a fool'. I'd like to see myself as a valid software engineer, I studied for it anyway, but I'm too impatient to go deep into the details of languages, libraries, frameworks, IDEs, etc. every day, so I ended up in another part of IT.

If having AI means a difference of having a tool like this and not having it, then I'm certainly for AI.

I have a project under my belt that I used LLMs a lot in, but the scope of the project revolves around what I can actually understand and what I've learned in those specific problems that I'm trying to solve. So I certainly don't see a problem with using LLMs for doing the research and churning out a proof of concept, then refining it to a serviceable and maintainable level once it proves its worth.

It also explains why my code is messy probably, I lack hours of coding. But these days I'm very happy with AI; if you ask it the right thing and keep the context for it valid, it'll work fine I'd say. I added an md file to the project for AI models and human developers about the code. However I agree the code is not very readeable now for a human, I'll try to fix that later too. But I think the code itself is not that messy, let me know your thoughts about it. Basically the code itself is generated 100% with AI, I just am architect, product owner, system manager and tester...

Neither am I a full-time backend developer, as I've likewise gone from a different job in IT after the backend development field felt too oversaturated, ripe for the taking by LLMs. In my project (won't share it unless asked, so as not to sound like I'm promoting stuff of my own), the possible gaps in maintainability, which I myself have tried to close over time (including through various LLMs that aren't tied to a specific vendor) basically occur only due to the lack of traction of the project online, as well as some more popular (but definitely less advanced) alternatives taking the spotlight.

I would claim that the main problem is when the LLMs are being used for commercial projects that involve real-life expenditures. I wouldn't trust some vibe-coded slop with my money. But if it's an unpaid hobby where you at least loosely know what you're doing (since you're already employed in IT, as opposed to the people that vibe-code and fake it till they make it), you know how it should sound like, and if it isn't completely unmaintainable, then it definitely has a place to exist. In reality, maintainable code is more of a myth. You don't have to fold under the pressure of a random stranger on the internet who, in an angry comment, orders you to refine the code to be maintainable for a project that you spend many, many weekends on with zero profit.

Naturally, if it gains traction, then there would obviously be a case for making the code more maintainable and generally slowing down to read the code and not pushing more than one commit per day, because the proof-of-concept phase is done - it would be time to refine that tool, which your regular LLM won't be able to accomplish as easily as it churned out a proof-of-concept. Thankfully, there's a clear end in sight for a project like this, so there won't be as much scope creep. Once the various bugs have been squashed and the floor has been scrubbed of what remains of them, the project would actually be done and finished.

[niekniek]

"In reality, maintainable code is more of a myth"; that's something I agree completely with. People are looking at AI to black and white I'd say. Context is everything. People also loose context when they don't work on the particular code anymore. AI looses it quicker but is also able to regain it quickly. The big picture is the problem, a human is needed there, people have to direct & test it in a proper way, my own experience is useful there. And you can help the model behaving correctly, check out the king prompt for example. This is a great article as well, or rather essay; https://every.to/p/after-automation. The human hamburger!

But back to the tool, the regression is fixed now. On github a bunch of artifacts is available now, I'll do release soon. Building a vgmrips package is now ridiculously easy; try scripts/fetch_snesmusic.py. Here's Gradius 3 and Lufia I did with that.