XTTS License After Shutdown

[fakerybakery]

Related to #3488

Hi, With the recent announcement that Coqui is shutting down, would you consider switching the license to a more permissive one, ie Apache 2.0 or MIT? The technology behind it is incredible and it would be great for the OSS community if the model became open sourced. Thank you!

[platform-kit]

Seconding this

[FrontierDK]

I also second this

[iprovalo]

+1

[dario248]

Seconding this

[erogol]

Currently we don't have plans to change the license. It'd also be unfair for people who paid for the license.

[slavakurilyak]

License remains MPL-2.0 (Mozilla Public License Version 2.0)?

[fakerybakery]

Hi @erogol, Thank you for your response. It makes sense that you are trying to keep licensing fair for existing purchasers. I have several questions:

1. Will new customers still be able to purchase a commercial license?
2. Who will own the IP of Coqui (& licensing rights) after the shutdown?
3. Have you considered keeping the CPML license until after all commercial licenses have expired (1 year from now) and after that switching to a more permissive license?

Thank you in advance.

[erogol]

1. most probably no
2. noone
3. I think there will be better models in that timeframe and maybe we train one

[fakerybakery]

Hi,

So, essentially, nobody new will be able to use it commercially, however Coqui will not earn any revenue from this limitation?

Also, for point 3, even if there are better models, personally, I still feel that XTTS will remain a valuable and important model for many years to come.

Thank you

[platform-kit]

@erogol I'm a paying customer, I bought the year long license just recently after the SaaS shutdown was announced.

As a paying customer I am sorry to say that my worst fear was confirmed with this latest announcement. I thought I was paying to use a software that would have a life for at least a year, not just a few weeks. Ideally I had hoped and expected that you guys would continue for many more years. That seemed to be the sentiment a month ago.

Again, as a paying customer (of both the Coqui Studio SaaS, and then the commercial XTTSv2 license), I am more miffed by the prospect of this becoming abandonware than I would be by the prospect of it becoming MIT licensed OSS.

Think about it. I've already invested not just money but man-hours into integrating the XTTS model into my product, which is now essentially rendered moot, since this product is going nowhere. So your decisions have actually created a waste/loss of my both time and money.

By contrast, if the product goes on to have a life as an MIT licensed model that the community adopts and maintains, then I have not incurred a loss of time or cash. Nor do I have to further spend money and time un-doing/re-doing the work of integrating a different model.

If you truly intend to abandon it from a legal point of view there is no difference to you between changing the license to MIT (or whatever equivalent open/highly permissible license), but there is a huge difference for us, the community, and specifically for those of us who paid for the commercial license. Open sourcing is the least you can do if you are not going to support it going forward.

I am very confident that if you asked a handful of your paying customers, they would ALL feel the same way.

Please consider open-sourcing. Option # 3 mentioned above (open-sourcing after 1 year has passed) might be the most courteous to your paying users, because at least then some kind of competitive advantage is conferred.

Thanks.

[fakerybakery]

Also, will paying customers who have already integrated the model into their code have to rewrite it because they cannot renew their license?

[q5sys]

Also, will paying customers who have already integrated the model into their code have to rewrite it because they cannot renew their license?

Effectively... yes. By not being willing to loosen the license for use after the license expires, Coqui is forcing everyone who built something around their platform to re-architect their pipelines/workflows/etc. It doesn't seem that they care much, due to the comment of "I think there will be better models in that timeframe...". It would seems that they expect people to be redeveloping their software anyway. It's unfortunate, but that's the result of businesses relying on software that has commercial restrictions through a license. It's always a 'Sword of Damocles' that someone else controls.

Everyone who paid for a license now has a 1 year ticking timebomb to find another solution or be in violation of their license. (Unless Coqui decides to change the license after that 1 year ends)

[harmony-ai-solutions]

The question in such cases of course is also, what's going on under the hood which we cannot see? Despite Coqui shutting down, there's probably still some money in the company and also Stakeholder interests in regards to their technology, the model, etc., and therefore they might have their hands tied on the model license, even if they wanted to change it to something more open.

However, from a client perspective, I'd second to what @platform-kit said. It's for sure certain that there will be better AI TTS tools in the future, but switching over to a new tech isn't super trivial, and even more complicated, if you've users who built custom voices based on a specific model. Since it's quite impossible to migrate them 1:1 and maintain their characteristics.

Also, I planned adding Coqiu Studio as a backend option into our software for example, so users who owned a Coqui-Studio subscription could have used your backend as another alternative, if they cannot (or didn't want to) run it locally. Actually I'm kinda glad I didn't start working on that yet. Because that shutdown really came out of nowhere.

Now that Coqui is shutting down, I considered to provide a hosted XTTS API for our users instead, since compute would be availiable for this. (Btw what happened / will happen to existing SaaS users now? Are they basically lost in the void now?)

However, this opens up legal questions with the current license, since despite our project is privately + community-funded, we would of course apply different rate limits to non-supporters using such an API, compared to users in higher tiers, which gives it some sort of commercial character after all.

[q5sys]

Despite Coqui shutting down, there's probably still some money in the company and also Stakeholder interests in regards to their technology, the model, etc., and therefore they might have their hands tied on the model license, even if they wanted to change it to something more open.

If we take what @erogol stated... that's not the case.

2. Who will own the IP of Coqui (& licensing rights) after the shutdown? 2. noone

However if he's being honest and truthful... there'd be no one to pursue someone using the purchased license after the 1 year term ends.

Actually I'm kinda glad I didn't start working on that yet. Because that shutdown really came out of nowhere. Agreed, I had some ideas, but they were pretty rough so I wasn't going to spend any money on it right now until I can flesh things out more. What makes this all more questionable to me though... is that the month prior they were being VERY active on Twitter to promote 1 year licenses. We don't know how many sold, it may have only been a few, but if my business bought a 1 year commercial license from another company, and then 30 days later they closed up shop and stopped providing services... I'd be speaking with my attorney. Having not bought a license, I don't know what the fine print on the terms were, it's possible the Dec 2023 license sale included some very fine print that was put in so they could just take everyone's money and run. (I hope that's not the case)

I wish the people at Coqui all the best in the future, and I'm sad they had to close up shop. But promoting people to buy a license and then rug pulling and then refusing to at least allow others to maintain the code for their own license that they just bought... just feels really wrong to me.

[pablodz]

To say only that many companies use really old open-source models like this one, over 4 years old and still in production. The fact that there are better and more efficient models doesn't mean that the community doesn't use the old ones, as mentioned above. It's a significant investment to switch from one framework to another, and companies prefer to bear the extra cost for a longer period until they are compelled to use the newer ones.

[FrontierDK]

This is true. I have helped a lot of companies, and to save cost, a lot of companies run outdated software - usually in virtual enviroments, where broken hardware isn't an issue. I still have a few clients who run win95/nt4 to control production equipment. So Coqui can, without problems, run in a virtual enviroment for decades.

[SomeDevWeb]

YES, PLEASE make it open-source!!! 🥺 Everybody would benefit, including Coqui...

OK, I have a clarifying question then. If I install the models on my personal PC and generate some voices that way, am I allowed to use them for a youtube video? Thanks!

[fakerybakery]

@SomeDevWeb

OK, I have a clarifying question then. If I install the models on my personal PC and generate some voices that way, am I allowed to use them for a youtube video? Thanks!

IANAL, but I'm pretty sure the license prohibits commercial use of any audio outputted from the model, even if it's on your own computer, unless you own a license. It would be super sad if people can't use this model any more, please consider open sourcing it @erogol @JRMeyer.

According to Coqui AI's FAQs,

Q. Can I sell content that I produced after the license finishes? Yes. If you make an audiobook today with XTTS, you can sell that audiobook even after your XTTS license expires.

Q. Is this licensing for unlimited web-based SaaS API usage or is the license for self-hosting / on-premise / locally running the model? It's a bring-your-own GPU deal. You can run XTTS on your servers, on a macbook, on a desktop computer... wherever you want. While you have an active license, they're your model weights, you run them where you want.

DISCLAIMER: I am not a lawyer. The above text is not legal advice.

[SomeDevWeb]

IANAL, but I'm pretty sure the license prohibits commercial use of any audio outputted from the model, even if it's on your own computer, unless you own a license.

Hi. Sorry, I kinda disagree, after what I saw here:

Q. Do I need the license if my use-case isn't commercial? Nope. If you're using XTTS non-commercially, it's free. You're already covered under the default license (the Coqui Public Model License).

And in that licence it says: Non-commercial Purpose Non-commercial purposes include any of the following uses of the model or its output, but only so far as you do not receive any direct or indirect payment arising from the use of the model or its output. Personal use for research, experiment, and testing for the benefit of public knowledge, personal study, private entertainment, hobby projects, amateur pursuits, or religious observance.

... which is exactly what one would do with a video. Of course, there's the consideration of the potential youtube monetization, which presumes the video is not producing any income. Only in the future if potentially the channel would become profitable, then in theory I suppose there will be a need to buy a licence, which obviously can't be done, so then I presume that's that...

But still, it's SUCH A BIG SHAME for this project to die. I really hope someone will continue it. In reality, I think all it needs is a fork of the FOSS repo and someone (or more people) to lead...

But who knows, we'll see what happens...

[q5sys]

In reality, I think all it needs is a fork of the FOSS repo and someone (or more people) to lead...

It's not the repo that's the issue. This repo and everything in it is under the MPL-2.0 license The issue is that the models which are on Huggingface are under the CPML. Anyone/Everyone can fork this repo and keep using XTTS under the MPL 2.0 license. But someone would have to pitch in and re-create the Models that Coqui made. We cant just retrain the model, because that's still using their original model and could be considered an output from that model. We need a custom freshly trained model with open weights.

Coqui could re-license the model, but they have indicated that they wont. Coqui could publicly release details of their initial training pipeline and information about their datasets, but I doubt they will.

They have stated that "no one" will own the IP or rights. So what value is there in keeping it private? Personal benefit. The knowledge of what they did and how has value... My guess is that the employees are hoping that their private knowledge of how it works will land them a job somewhere, so they are incentivized -not- to share anything that they don't have to. If they truly believed there's no value to any of it... then there would be no reason to let it be lost.

The shutdown coming so soon after they went on a social media campaign in Dec of 2023 selling 1 year licenses really rubs me the wrong way, and I cant help but think there's something else at play. Maybe there's not, maybe they really did try to get people to buy something they knew they would kill off one month later. But with silence about the reasons and not being willing to open source the models, I cant help but assume that there's details that are private that explain why everything has gone the way it has.

[metal3d]

Currently we don't have plans to change the license. It'd also be unfair for people who paid for the license.

It's unfair to leave some people having a "for life" right to freely use a model, and to not authorize new users to have the same rights.

XTTS model is fantastic, and I'm sad that the society couldn't continue. Really. But, now... it's too late.

Think about the Blender foundation. Blender was proprietary but "freely usable" until NaN company shut down. Ton Roosendaal made a hard work to make Blender open source and to free to use. Just take a look on it now...

Restriction is not good at all, it only makes the project to be unusable in the real world (the world where we need money, yep).

Leaving the XTTS model freely usable could help donators to give a coin.

[fakerybakery]

+1 @metal3d

It's unfair to leave some people having a "for life" right to freely use a model, and to not authorize new users to have the same rights.

I don’t think anyone has a license for life, the licenses were all 1 year, right? Unless there were licenses issued before that

[metal3d]

+1 @metal3d

It's unfair to leave some people having a "for life" right to freely use a model, and to not authorize new users to have the same rights.

I don’t think anyone has a license for life, the licenses were all 1 year, right? Unless there were licenses issued before that

Yes that's right.

But what happens after one year for those who are based their business on this model?

And there is a weird mention about the model "in the form provided by the licensor". What is this form? Is a fine tuned model not in the form provided by the licensor, and so, not covered by the CMPL?

The license seems nebulous and lacking in detail. It only provides restrictions and not authorizations in the case of fine tuning, and above all it does not indicate the conditions in the case of purchase (which is no longer possible) of commercial exploitation rights.

I'm not a lawyer, but it seems to me that there are gaps and that I could hardly be asked for trouble if I train my model from the model supplied, even in the case of fine tuning.

Of course, the xtts owner will say that I don't have the right to use it for my futur monetized YouTube video... It's a pity. Xtts is impressive, I fine tuned the model with my dataset and the generated voice is near the perfection. My wife didn't find the difference between my voice and the generated one.

[fakerybakery]

I think the only option left is to re-train XTTS since the code is MPL licensed, but that would be very expensive

[harmony-ai-solutions]

I think the only option left is to re-train XTTS since the code is MPL licensed, but that would be very expensive

We might be able to give that a try when we finished training our new one-shot TTS; but we'll still be busy with that for some time.

However I didn't check the license in detail; is it possible to use custom trained weights trained from scratch for this model in commercial context, too ?

[metal3d]

That my question about the "form provided by the licensor" in the CMPL license. A fine tuned model isn't the model in the same form, and it's not an output of the model. It's a derived model.

The CMPL is absolutely unclear on this point, and doesn't even mention it. If I were a little dishonest, I'd say that my model is not the one provided by CoquiTTS, and that I therefore have no restrictions on using it as I see fit.

In fact, I'm sure this goes against their desire to restrict commercial use of the XTTS V2 model. And that makes me a little bitter. I've always found this kind of restriction terribly harmful. The only way for CoquiTTS to be revived commercially is for people to use it and ask for support.

In my opinion, this is where CoquiTTS went wrong. They wanted to sell access to the studio and sell the model... Many companies have tried this model and failed.

I'll say it again, but that's what happened with Blender, and it was a disaster for the publisher. What saved Blender was precisely that it gave everything to the community. The money came from donations and the sale of support, but not from the product.

Hybrid open source is almost always doomed to failure. You either have to go full-opensource, or full-ownership (and have some pretty strong arguments in favor of this mode of operation).

So, to sum up: a fine-tuned model is not mentioned in the license, we don't have an answer on the subject, and it seems imperative that the CoquiTTS managers rid us of these absurd clauses since the box has crashed.

[fakerybakery]

I think the only option left is to re-train XTTS since the code is MPL licensed, but that would be very expensive

We might be able to give that a try when we finished training our new one-shot TTS; but we'll still be busy with that for some time.

However I didn't check the license in detail; is it possible to use custom trained weights trained from scratch for this model in commercial context, too ?

The code itself is MPL licensed, only the weights are licensed under CPML. So if you completely retrained everything from scratch you probably could (but disclaimer IANAL), finetunes would probably still fall under the license. But I might be wrong

[louispaulet]

IANAL, but retraining a model for legal reasons is ecologically wrong.

If you do not change the license for future hobbyists, at least do it for the planet.

[choilive]

IANAL, but retraining a model for legal reasons is ecologically wrong.

If you do not change the license for future hobbyists, at least do it for the planet.

Eh, ecologically wrong? How so? Training cost? Training a model costs basically nothing in terms of energy. The internet uses about 2% of global energy, maybe we should shut that down too.

[Dmmoim]

We won't see them change the license since it is now part of Rabbit r1. Its co-founder works there. :(

[platform-kit]

Lame. They've been outclassed by other models anyway, but still, this is lame.

[jneuendorf]

Lame. They've been outclassed by other models anyway, but still, this is lame.

Do you have any reference to better (open-source) alternatives?

[platform-kit]

@jneuendorf https://github.com/metavoiceio/metavoice-src

1.2B Parameters vs XTTS's 750m.

[pablodz]

Forked :D https://github.com/idiap/coqui-ai-TTS

[eginhard]

Forked :D https://github.com/idiap/coqui-ai-TTS

Just to clarify that this fork doesn't change the license of the code (MPL 2.0) or the license of any pretrained models, including XTTS (CPML), so it's not really relevant to this issue.

[lukaLLM]

Any updates regarding it or what alternative do you use this looked like the best low-latency commercial solution. I go into dependency searching for tool hell to replace it