An AI-driven violin library - might be interesting to take a look

[joe_acestudio]

Hey all,

I'm Joe Guo, the founder of ACE Studio.

I'd love to introduce ACE Studio's newest update - an AI-driven violin library.

Here's a quick comparison between a traditional sample-based violin library, and ACE's AI violin, with only melody input:



Currently, it's a beta. There's tons of space to improve. And we'll add articulation and expression controls very soon. The idea of the AI violin is to generate a convincing performance with only MIDI input, but remain the possibility for people to tweak on the nuances.

Just to be clear, we're not trying to discredit the sample library, or overlook how amazing the performance it can achieve with decent programming. But the idea of this comparison is to showcase a quick scenario with only MIDI input - no CC controls, no keyswitching or whatnot.

Our aim is to offer a different kind of tool: one that makes expressive playing more accessible and effortless, especially for creators who don’t have the time or technical know-how to deeply program their instruments.

Would love to hear your thoughts.

Cheers!

[mcbpete]

At 0:40 the video says 'AI violin renders performances in the cloud' - so what's happening at the cloud end, software wise ?

[joe_acestudio]

At 0:40 the video says 'AI violin renders performances in the cloud' - so what's happening at the cloud end, software wise ?

The violin library, which is actually a neural-net model, is deployed on the server side. It takes the midi sent by the client as input, run the inference pipeline to generate the output, which is the violin sound, and send it back to the client side.

[MrJubbly]

How would this compare or differ to say, a physically modelled violin from Audio Modeling's SWAM instruments?

[Super Piano Hater 64]

The violin library, which is actually a neural-net model, is deployed on the server side. It takes the midi sent by the client as input, run the inference pipeline to generate the output, which is the violin sound, and send it back to the client side.

So, even though you call it a library, there's no way to use that part locally on the user's own hardware, because what the user actually interacts with is an always-online cloud service for which the "library" is an opaque implementation detail.

Do I understand this correctly?

[joe_acestudio]

How would this compare or differ to say, a physically modelled violin from Audio Modeling's SWAM instruments?

SWAM is great. But the aim of ACE's AI violin is different.

The advantage of SWAM is that you don't need to download gigabytes of a sample library. But you still need either to program or to play in real time along with all the CC controls, keyswitches and so on.

With ACE's AI violin, by only inputting MIDI, you can get a convincing performance.

Not saying SWAM is not good. But for people who don't have the know-how or the time to tweak with all the programming of a sample library, and who would like to focus more on the creative vision. ACE's violin is a good option.

[joe_acestudio]

The violin library, which is actually a neural-net model, is deployed on the server side. It takes the midi sent by the client as input, run the inference pipeline to generate the output, which is the violin sound, and send it back to the client side.

So, even though you call it a library, there's no way to use that part locally on the user's own hardware, because what the user actually interacts with is an always-online cloud service for which the "library" is an opaque implementation detail.

Do I understand this correctly?

Correct. And you can use this AI violin across different computers without needing a powerful machine, since all the heavy processing happens in the cloud.

[v1md]

Is there a way to control the performance parameters besides what the AI decides by default?

[_al_]

Hmm, your violins with AI articulations VS sampled violins with NO articulation...

The sampled violins are always gonna sound like crap with zero articulations, so it doesn't sound like a very fair comparison tbh.

I'm not knocking your ACE version... it does look interesting (and sounds pretty good too)

[IvyBirds]

Your demon sounds like you are using an early 1990s era Rompler to be the example of a "sample", and even then it sounds like you manipulated the envelope to have it sound grating on the attack phase

I have multiple sampled string libraries that would sound dramatically better than your example and I don't even have high end ones

Roland's "Super Natural" acoustic engine powering violins in the Integra 7 sounds significantly better also when it comes to articulations

Your demo video sounds good, but I have to question the accuracy of it. Since you obviously went out of your way to post the absolute worst example of what a sampled violin sounds like in 2025, what was done to make the AI version sound better

[joe_acestudio]

Is there a way to control the performance parameters besides what the AI decides by default?

As for the beta version that we just released several days ago, it only take melody as input and AI decides the articulation by default. But we'll include articulation and expression controls very soon in the near future updates. Our plan is to make the AI-driven virtual instruments somehow imitate what human artists are doing -- without articulation controls it will also perform convincingly realistic and expressive by default, but it will also listen to and align with the articulation and expression controls given by the user.

[joe_acestudio]

Your demon sounds like you are using an early 1990s era Rompler to be the example of a "sample", and even then it sounds like you manipulated the envelope to have it sound grating on the attack phase

I have multiple sampled string libraries that would sound dramatically better than your example and I don't even have high end ones

Roland's "Super Natural" acoustic engine powering violins in the Integra 7 sounds significantly better also when it comes to articulations

Your demo video sounds good, but I have to question the accuracy of it. Since you obviously went out of your way to post the absolute worst example of what a sampled violin sounds like in 2025, what was done to make the AI version sound better

Thanks for the feedback. But we really didn't intentionally make the sample library sound worse. On the contrary, we don't want any controversy. We want the video looks reasonable.

But the whole idea is to input melody only on both sides, to highlight the out-of-box performance of ACE's AI violin. (And I think it's the biggest differentiator for people to get the point why the world need a AI-driven violin VI if the sample-based ones are already so good)

That said, there's really not too much we can do (at least on the libraries we have) to make the sample-based one sound better if the restrict is inputting melody only -- no CC controls, no keyswitches or whatnot

We initially used CSS because several of our team members use it a lot. But it performed not well with just melody input. We switched to Chris Hein Solo String and Chris Hein Ensemble String, the Solo one performs good on lyrical pieces, but really bad on energetic pieces. And the ensemble one performs equally good for all pieces. So we finally picked the Chris Hein Ensemble String as the library to compare.

And we didn't do any post-process to make the sample library sound worse, or make the AI violin sound better. All the performances in the video can be reproduced. I can send you the Logic Pro X project file of the sample violin track as well as the ACE project file of the AI violin track if necessary.

We are totally aware of how good a sample-based violin can be if programmed decently. But the idea is to show that an AI-driven violin is a different kind of tool which can perform realistic and expressive sound even without any other inputs other than melody.

[joe_acestudio]

Hmm, your violins with AI articulations VS sampled violins with NO articulation...

The sampled violins are always gonna sound like crap with zero articulations, so it doesn't sound like a very fair comparison tbh.

I'm not knocking your ACE version... it does look interesting (and sounds pretty good too)

Thanks for the feedback. I agree as for the video we failed on conveying the idea

[joe_acestudio]

Hmm, your violins with AI articulations VS sampled violins with NO articulation...

The sampled violins are always gonna sound like crap with zero articulations, so it doesn't sound like a very fair comparison tbh.

I'm not knocking your ACE version... it does look interesting (and sounds pretty good too)

Thanks for the feedback. We really messed up on conveying the key point behind this comparison.
We didn't meant to discredit the sample-based libraries. And we 100% know how good they are if programmed well. But the idea of this video is to showcase the performance if the only input is melody - no CC controls, no keyswitch articulations, etc.

We think melody-input-only is quite important for this video. It's not about to make our AI violin sound BETTER, but to make the point that ACE's AI violin is a DIFFERENT tool -- one that makes expressive playing more accessible and effortless, especially for people who don’t have the time or technical know-how to deeply program their instruments.

" If we'd introduce a new kind of tool to the musician world, we'd like to strongly highlight the differentiator of it. Otherwise why the world need an AI violin if the sample ones are already so good" -- that's our initial thought. But I agree the video narrative has a space to improve. Apologise for mis-conveying the point.

[mambo888]

I'm not seeing the computational power required for this to be in the cloud. This should just be an add on / macro or midi FX either built into a sample library or as a plugin that adds articulation to the midi after analysing the track. Personally, I'm not interested in anything that cannot be used offline, however I think this is a good market opportunity to develop an AI midi FX plugin that analyses your track and the based on the style of your piece adds the relevant articulations either real time or after analysing the midi.