A couple of cases where A.I. doesn't suck

[TimC340]

I've used AI quite a lot in trying to work out techniques in what has been a pretty intensive music-tech-learning year for me. And, as VP implies, I've found that AI is often wrong - convincingly, but wrong nonetheless. Indeed, just go and ask AI about some of the techniques used in Studio Pro and I guarantee you that you'll be shouting at it within a few minutes - it'll dig up stuff from Studio One v3 and present it as the correct technique for FSP8, for example. I've lost count of the number of times I've told Gemini/Google AI to restrict its answers to only facts relevant to FSP8 - and it still doesn't get it. (I hope that the AI help bot in the Plus subscription does a better job, but I don't have that!).

I've no doubt that more advanced AI models can be guided and constrained to be more accurate, but changes announced in the last week or so to the AI access available to general domestic Google users suggests to me that the enshittification of AI access has started and is unlikely to improve. Basically, if you don't pay (a lot) for it, don't expect much from it.

 

[Vocalpoint]

But changes announced in the last week or so to the AI access available to general domestic Google users suggests to me that the enshittification of AI access has started and is unlikely to improve. Basically, if you don't pay (a lot) for it, don't expect much from it.

This started long ago.

Still amazes me how many references I see in forums now where the post starts "Hey guys - I ran this through Gemini..." or "Hi - I have this issue and ChatGPT says..."

And in the end - after all their Ai slop BS and back and forth - the chatbot ends up knows nothing about their actual issue and more often or not - it comes down to us humans to save the day for the poster.

If I listen closely - I can hear perfectly healthy brains shutting down across the internet....

VP

 

[TimC340]

Since I'm in the final stages of my current music production, this topic has been floating around in my head.

I remembered reading something different about how Synthesizer V works, but I didn't want to spend too much time researching it myself while working on my music. So I asked an artificial neural network to look into it and check the technical details.

Here's what it found:
Dreamtonics actually used a hybrid sample-based and neural approach in earlier versions of Synthesizer V. However, the company introduced fully neural singing synthesis in 2020, and Synthesizer V Studio 2 uses a rewritten neural engine.

So the description isn't entirely without historical basis, but it doesn't accurately describe the current technology.

Source: https://www.dreamtonics.com/how-we-...nd-even-better-in-synthesizer-v-studio-2-pro/

The voices in Synth V Pro v2 are still derived from real people, who are sometimes identified in the blurb on Dreamtronics' website. The mechanics of the exact algorithms used to reconstruct the voices in the program may have changed since V1 (and not necessarily for the better in some cases), but it still involves sampling those voices as Job 1.

 

[TimC340]

This started long ago.

Still amazes me how many references I see in forums now where the post starts "Hey guys - I ran this through Gemini..." or "Hi - I have this issue and ChatGPT says..."

And in the end - after all their Ai slop BS and back and forth - the chatbot ends up knows nothing about their actual issue and more often or not - it comes down to us humans to save the day for the poster.

If I listen closely - I can hear perfectly healthy brains shutting down across the internet....

VP

To be fair, I have found the AI chatbots to be quite useful on subjects where a huge amount of background info is available and accessible. Recuperation after a specific operation, fitness techniques, the finer details of Garmin watches are some of those that have worked well for me. But for specialist and relatively obscure (not to mention subjective) topics like music production, it's less good, shall we say! The trouble is that a lot of the online source information is contradictory and often depends on the musical tastes and opinions of someone who was perhaps a music or music tech journalist for a short time before moving on to something else. In the days before AI, when searching was down to the individual, I'm sure we all came across these issues and AI is simply reflecting that - it just brings up the contradictory crap quicker than we could! It was just easier to judge the validity of a search result yourself than it is to argue with something that presents 'facts' convincingly...

 

[FinalGod]

The voices in Synth V Pro v2 are still derived from real people, who are sometimes identified in the blurb on Dreamtronics' website. The mechanics of the exact algorithms used to reconstruct the voices in the program may have changed since V1 (and not necessarily for the better in some cases), but it still involves sampling those voices as Job 1.

Yes, absolutely—the voices originate from real singers. I wasn’t disputing that, and it’s entirely consistent with what I posted.
The distinction is between recording a singer to train a neural model and assembling recorded vocal samples during synthesis. Both start with human recordings, but they are different ways of producing the resulting vocal.
Dreamtonics explicitly describes the early engine as selecting and crossfading samples, followed by fully neural synthesis and a rewritten engine for Studio 2. Their explanation is here.
So my correction concerned the original claim that the AI merely joins syllables and phonemes together—not where the voices came from. Whether someone prefers the older sound is a separate, perfectly valid preference.

 

[gabo]

One good thing about AI...

It is teaching us to QUESTION answers! Which applies not only to AI, but to answers we get from humans. I have spent most of my career questioning things and doing actual experiments to find the truth.

AI doesn't give correct answers because the info it gets to construct the answers comes from humans... who are quite often wrong!

Hopefully AI will continue to encourage humans to question everything, only come to a conclusion about the answer after doing things and experiencing the actual results. And even then, continue to question it as maybe your experiment is flawed or your understanding of the results are only partially correct. Be a "dog with a bone" and never give up searching for the truth!

 

[FinalGod]

You don't. You do you. No one is forcing you to actually learn anything.



That only works if you can get past that gigantic assumption you are making that the AI is correct. Only you can teach you how to do it correct.


FWIW - I already use pen and paper and plenty of "modern tools" and have taken the "unnecessary" time to master their technical disciplines for my workflow.

If you do not wish to invest any time to really learn your toolkit and instead favor shortcuts due to inability, laziness or whatever LLM is a hot excuse today - that's your call.

VP

I think you're making quite a few assumptions about me here.
Where exactly did I say that AI is always correct? I never claimed that, nor did I suggest that its answers should be accepted without question.
I said AI can help people learn. Just like books, tutorials, teachers, or other resources, the information can be wrong. That doesn't make the resource useless.
And choosing not to master every technical discipline involved in a project doesn't mean I'm unwilling or unable to learn. It simply means I decide where I want to invest my time.
If you enjoy learning every technical detail of your tools, that's perfectly fine. I have no problem with that.
But why should choosing a different approach automatically be considered laziness or inability?
There's no need to make it personal or imply that someone is lazy or incapable simply because they choose a different approach.
For example, I recently used Codex to create a small Windows utility that extracts plugin information from Studio One project files. I defined what I wanted, tested the results, identified limitations, and learned a few things about the .song file format along the way.

I could have spent weeks learning C# and Windows Forms to write it myself. But that wasn't my goal. My goal was to have a useful tool.

And now I have one.
That's exactly what I meant by choosing how to use technology.

 

[tulamide]

Since I'm in the final stages of my current music production, this topic has been floating around in my head.

I remembered reading something different about how Synthesizer V works, but I didn't want to spend too much time researching it myself while working on my music. So I asked an artificial neural network to look into it and check the technical details.

Here's what it found:
Dreamtonics actually used a hybrid sample-based and neural approach in earlier versions of Synthesizer V. However, the company introduced fully neural singing synthesis in 2020, and Synthesizer V Studio 2 uses a rewritten neural engine.

So the description isn't entirely without historical basis, but it doesn't accurately describe the current technology.

Source: https://www.dreamtonics.com/how-we-...nd-even-better-in-synthesizer-v-studio-2-pro/

You misinterpret the sales promotion. It is important that you understand, why a difference is made between a database and a neural network in the first place. A database contains data with indexes that allows external algorithms to access the data. A neural network is a net of mathematical numbers, of propabilities that stand for the data. Or in short: the data is part of the network, but not in form of a database, but as propability clouds (like banana, yellow, bent are one cloud, but citrus, yellow, sour as well).
This doesn't change the nature of the singers. They still are a colletion of consonants and vowels recorded from professional singers. They still aren't generated from thin air, but taken from those samples. But they don't exist as databases, where one neural network would control any of them. They are now their own neural network each (simplified to make the point clearer).
So, if your concern is wether the consonants and vowels are real or synthetic - they are real. If your concern is, how they are stored and interconnected - they are the vocabular of the neural network, not their data.

 

[SwitchBack]

And who says a human can't make mistakes just because they're an "expert"?
Everyone makes mistakes. Humans, animals, computers.
The real question is whether we should blindly trust any of them.
If a GPS tells you to turn left and you drive straight into a stone wall, two mistakes have been made. The computer gave incorrect directions, and the human blindly followed them without checking what was actually in front of them.
And just because a dog barks doesn't necessarily mean someone is standing at the front door.
But that wasn't my point in the first place.
The question of whether we should blindly trust computers has been debated for decades. But that's not what I was talking about.
My point was simply that AI is a tool, and it's up to us how we choose to use it. It can even help us learn things ourselves rather than just doing them for us.
It can help us get things done faster, create things, or even make them possible in the first place.
Take the example from the thread starter — he used AI to help him with a task. As far as I know, he survived the experience...

This isn't even about who's right and who's wrong, but about who you trust when AI and human expert are at odds, especially when you yourself are not an expert. When you say "I use AI to teach/train me..." you state as much.

Some in my family have a medical profession. They tell me that ChatGPT has become a serious problem because people refuse treatment because 'the computer' gave them a different diagnose. The fact that doctors have diplomas and AI does not could not convince them otherwise. Humanity is going through a learning curve regarding the use of AI, we're in our infancy if you like. And yes, it's a tool but so is a box of matches and we don't hand those out to kids, not without proper supervision.

 

[FinalGod]

You misinterpret the sales promotion. It is important that you understand, why a difference is made between a database and a neural network in the first place. A database contains data with indexes that allows external algorithms to access the data. A neural network is a net of mathematical numbers, of propabilities that stand for the data. Or in short: the data is part of the network, but not in form of a database, but as propability clouds (like banana, yellow, bent are one cloud, but citrus, yellow, sour as well).
This doesn't change the nature of the singers. They still are a colletion of consonants and vowels recorded from professional singers. They still aren't generated from thin air, but taken from those samples. But they don't exist as databases, where one neural network would control any of them. They are now their own neural network each (simplified to make the point clearer).
So, if your concern is wether the consonants and vowels are real or synthetic - they are real. If your concern is, how they are stored and interconnected - they are the vocabular of the neural network, not their data.

Thanks for the explanation! I think we may have been talking past each other a little.

I never claimed that the voices were stored in a database. My point was actually the opposite: Dreamtonics describes how Synthesizer V moved away from its original sample-selection and crossfading approach toward fully neural singing synthesis.
Of course, the voices are based on recordings of real singers. That was never in question. The distinction I was making is between using recordings to train a neural model and selecting and joining recorded audio fragments during synthesis.
For an independently documented example, there's also the scientific paper NNSVS: A Neural Network-Based Singing Voice Synthesis Toolkit. It describes how neural models can be used to predict acoustic features and generate singing audio.
This doesn't establish the exact architecture used by Synthesizer V, but it illustrates the technical distinction I was referring to.
So I suspect we're actually in agreement on the main point: human recordings provide the training material, while the trained neural model is responsible for synthesizing the resulting voice.

 

[FinalGod]

Some in my family have a medical profession. They tell me that ChatGPT has become a serious problem because people refuse treatment because 'the computer' gave them a different diagnose. The fact that doctors have diplomas and AI does not could not convince them otherwise. Humanity is going through a learning curve regarding the use of AI, we're in our infancy if you like. And yes, it's a tool but so is a box of matches and we don't hand those out to kids, not without proper supervision.

I completely agree with you on that!

That's not how AI should be used.

 

[Vocalpoint]

But why should choosing a different approach automatically be considered laziness or inability?

Please excuse my skepticism here but any "approach" where the answers to any problem are simply given to you -with no effort whatsoever - does have a name.

I will let you decide which of the ones you mentioned above - apply.

VP

 

[FinalGod]

A different approach? Like one where you think for yourself?

Your "approach" it seems - is to have AI give the answers as quickly as possible with no thought at all.\

AKA - do it all for you.

VP

I'm going to end this discussion with you here.
You keep implying that people are lazy or incapable of thinking for themselves simply because they choose to use AI tools.
If you don't want to use AI, that's perfectly fine with me. I respect your choice. It would be nice if you could show the same respect for other people's choices instead of insulting them for it.
There's no need to derail the discussion or make personal remarks about people simply because they choose a different approach.
And to me, that's not what music is about. Music should be accessible to everyone, regardless of their technical skills or the tools they choose to use. That's what matters to me.
Have a nice day.

 

[Vocalpoint]

There's no need to derail the discussion or make personal remarks about people simply because they choose a different approach.

No one is promoting a derail.

If this is how you are taking it - then please remember that discussions in here are both measured and public.

If this thread is suddenly off limits to anyone but those who believe that Ai is required to do their heavy lifting for them - then I would appreciate if the mods labeled it as such.

No one said you need to like the a discussion just because someone disagrees with your approach. But it will happen to all of us here eventually.

You have a great day too.

VP

 

[TimC340]

My approach to getting tools is to find what I think I need by recommendation or review and buy it. If it doesn't exist yet, I move on. In my eyes that's far lazier than using Claude or whatever to develop a tool that does exactly what the user needs at the time they need it. I have absolutely no problem with that approach, but it's not one I'd take myself simply because the process itself would take more time than I'd want to give it. That's why I didn't develop my own DAW, or EQ, or reverb. I do use AI to do some research for me, but I check its results carefully (often it'll cite YouTube videos that were panned by commenters at the time, for example!). That just tells me that the depth that publicly and freely available AI goes to in its research is very shallow, and thus often wrong. However, I'm aware that better models exist and are accessible (for a price), and that interesting things are happening as a result of those models.

 

[FinalGod]

I do use AI to do some research for me, but I check its results carefully (often it'll cite YouTube videos that were panned by commenters at the time, for example!). That just tells me that the depth that publicly and freely available AI goes to in its research is very shallow, and thus often wrong. However, I'm aware that better models exist and are accessible (for a price), and that interesting things are happening as a result of those models.

I agree with u.

I think there's also a common misconception that AI systems like Claude, Gemini or ChatGPT always have access to the latest information.
They may have tools to search the internet and retrieve current information, but their underlying training data can be months or even longer out of date.
And since these are general-purpose AI systems, they won't necessarily have accurate information about every specific piece of software, especially when newer versions have changed how things work.
One workaround I've found useful is to give the AI the current user manual for the software I'm working with. That gives it a much better reference to work from.
Of course, that doesn't mean it can't misunderstand the documentation or make mistakes. AI isn't perfect. But it can still be a very useful tool for solving problems.

Here's a real example from my own experience.

I spent about two weeks trying to get the MCP interface in DaVinci Resolve Studio working. I was going back and forth with Blackmagic Design support, trying different things, but eventually felt we weren't making much progress.
As a last idea, I decided to let Codex investigate the problem directly.
It analyzed the environment, identified the underlying issue, developed a workaround, and helped resolve the actual problem.
That took roughly 15 minutes.
After two weeks of troubleshooting, that was quite an experience.
And yes, I had access to those capabilities through my ChatGPT Plus subscription.

My point isn't that AI always knows better than a human expert. It clearly doesn't. But when used appropriately, it can provide another way to investigate problems, learn something new, or accomplish tasks that would otherwise be difficult or time-consuming.

 

[tulamide]

Thanks for the explanation! I think we may have been talking past each other a little.

I never claimed that the voices were stored in a database. My point was actually the opposite: Dreamtonics describes how Synthesizer V moved away from its original sample-selection and crossfading approach toward fully neural singing synthesis.
Of course, the voices are based on recordings of real singers. That was never in question. The distinction I was making is between using recordings to train a neural model and selecting and joining recorded audio fragments during synthesis.
For an independently documented example, there's also the scientific paper NNSVS: A Neural Network-Based Singing Voice Synthesis Toolkit. It describes how neural models can be used to predict acoustic features and generate singing audio.
This doesn't establish the exact architecture used by Synthesizer V, but it illustrates the technical distinction I was referring to.
So I suspect we're actually in agreement on the main point: human recordings provide the training material, while the trained neural model is responsible for synthesizing the resulting voice.

We are! I'm sorry that I misunderstood your point. Yes, the neural network is especially useful to define edge cases, where a database would be out of luck. For example, the "sh" sound will start sligthly different depending on the last letter before it. "ashamed" versus "full of shame", for example. You can't record any possible difference in initial pronounciation of that sound. But a neural network can identify the closest blend of sounds from the vocabulary and make it sound much more natural. And that was only possible, after they switched from database accessed by neural network to full neural network. Of course, that is only one of many examples.