Real-Time-Voice-Cloning pretrained.pt (Bluefish models link ) seems missing?

Hi, I am testing / running this code in a class project and noticed today that the pretrained.pt file in synth, vocoder and encoder (previously stored in /saved_models and /saved_models/pretrained/ ) seems missing?

Asked Jan 09 '22 11:01
avatar RobbeW
RobbeW

15 Answer:

Aha, I guess issue https://github.com/issues/CorentinJ/Real-Time-Voice-Cloning/928 answers my concern, but kinda does not solve it. (ghost made me miss that this comment was from BlueFish).

Maybe on the off chance anybody (@sveneschlbeck ) still has some pretrained.pt's stored locally?

1
Answered Dec 01 '21 at 14:48
avatar  of RobbeW
RobbeW

@RobbeW I do, this is a hell of a lucky day we're having. I coincidentally downloaded the repo two days back or so but were not able to do something yet. This needs to be fixed, otherwise the link in the README is leading to an error page, too.

The files are too big for an upload, I'll try to add them in a new commit, will keep you posted

@CorentinJ What are you proposing on doing here?

1
Answered Dec 01 '21 at 19:02
avatar  of sveneschlbeck
sveneschlbeck

@RobbeW @CorentinJ We got a problem, the files are way too large to just upload them and Git LFS is not working on forked repos, so we really depend on @CorentinJ to add or restore them...

1
Answered Dec 01 '21 at 19:21
avatar  of sveneschlbeck
sveneschlbeck

Hello, @RobbeW! I also have the files, I have been trying to train a spanish model for a month, with no luck. I'm also doing a class project. I can pass you the files via gmail or drive as a fast solution for you if you need them fast. By the way, I need some help with the work of training in spanish, since I'm running out of time and the datasets I have worked with haven't given me results (#789).

1
Answered Dec 01 '21 at 19:24
avatar  of AlexSteveChungAlvarez
AlexSteveChungAlvarez

@RobbeW In the docs, there's still an initial commit .zip file: https://drive.google.com/file/d/1n1sPXvT34yXFLT47QZA6FIRGrwMeSsZc/view Should contain all three pretrains according to the file size

1
Answered Dec 01 '21 at 19:25
avatar  of sveneschlbeck
sveneschlbeck

That should work for him.

@RobbeW In the docs, there's still an initial commit .zip file: https://drive.google.com/file/d/1n1sPXvT34yXFLT47QZA6FIRGrwMeSsZc/view Should contain all three pretrains according to the file size

1
Answered Dec 01 '21 at 19:27
avatar  of AlexSteveChungAlvarez
AlexSteveChungAlvarez

@RobbeW @AlexSteveChungAlvarez This is a pretty good example of why we should definitely improve the docs on this repo. There were several people wishing tutorials on trainings (also in other languages) and deleting pretrained models/accounts without notification is something that should NEVER happen. Don't know what went wrong there...

Should wait for @CorentinJ `s opinion

1
Answered Dec 01 '21 at 19:28
avatar  of sveneschlbeck
sveneschlbeck

@sveneschlbeck @AlexSteveChungAlvarez thanks a lot for the help. Kinda also my own fault, since I used the Colab version of the code containing paths to the three pretrained.pt files, which relied a !wget and the model blue-fish contributed. I should have, at least once, bothered to make a local copy of the file tree.

I will give it a go with the G-Drive link above. Will try and figure out how to adjust the code to use the synthesizers multiple files, instead of the pretrained.pt single file path reference.

If I can contribute in any way, let me know. If some solution comes out of this for my classes and students, I owe you all a coffee / beer.

1
Answered Dec 01 '21 at 19:47
avatar  of RobbeW
RobbeW

Can any of you help me with the spanish model for the synthesizer? Issue https://github.com/issues/CorentinJ/Real-Time-Voice-Cloning/789 has my work until last week, but since then I just tried with subsamples of the last dataset I used for training. It hasn't given me good results, though. I think I will need to find another spanish dataset. The main problem is that I just have 1 month left to achieve it and each try has taken me about 1 week.

1
Answered Dec 01 '21 at 20:12
avatar  of AlexSteveChungAlvarez
AlexSteveChungAlvarez

@Tymoteusz pointed out this does not contain the synthetizer pretrained model, but I don't know why the comment is not here anymore. I read it and uploaded the 3 pretrained models to this drive: https://drive.google.com/drive/folders/1lb-LlS8Sx9RqcGzuV6GxvKHk-PC9TqQx?usp=sharing . It is not a zip, so anyone who needs the models may download each of them. They are ordered in the original structure.

@RobbeW In the docs, there's still an initial commit .zip file: https://drive.google.com/file/d/1n1sPXvT34yXFLT47QZA6FIRGrwMeSsZc/view Should contain all three pretrains according to the file size

1
Answered Dec 01 '21 at 23:44
avatar  of AlexSteveChungAlvarez
AlexSteveChungAlvarez

@AlexSteveChungAlvarez Apparently the archive mentioned earlier contains older models which are not compatible with current master branch (source).

I have tested your pretrained models and it works without any problems 👍

1
Answered Dec 02 '21 at 00:23
avatar  of delyro
delyro

Tested the pretrained.pt models and they work! Marvelous and a big thanks! I'll store them locally this time and have a version in a cloud service so my students can !wget the files (using OneDrive) If Sven or CoretinJ would deem it necessary to store critical files on a more secure cloud storage, I'm willing to chip in.

1
Answered Dec 02 '21 at 13:03
avatar  of RobbeW
RobbeW

@AlexSteveChungAlvarez Sorry I couldn't help you with the Spanish model last week, I don't have the resources or skill to do so, but maybe this new paper might help you reach that goal and deadline (TTS language synthesis, zero shot, low resource languages ...)

1
Answered Dec 07 '21 at 08:42
avatar  of RobbeW
RobbeW

This paper seems very interesting, I already had a quick look at Coqui TTS, which is the tool they used for their experiments (Coqui TTS was developed by the ex team mozilla). Hopefully I cant take a look at it later. Maybe you can give me a hand in understanding Coqui? I think you should close this issue since the issue was solved already, and we should talk about the helping hand in https://github.com/issues/CorentinJ/Real-Time-Voice-Cloning/789 :) .

1
Answered Dec 07 '21 at 13:44
avatar  of AlexSteveChungAlvarez
AlexSteveChungAlvarez

@AlexSteveChungAlvarez could you check on your local machine when the files were created? We are trying to figure out over here https://github.com/issues/CorentinJ/Real-Time-Voice-Cloning/942 if those from 06.02.21 are latest, thus RTVC 7. If your Date is newer, then we are behind. If I download unzipped files from your drive they get date of today.

1
Answered Dec 09 '21 at 20:48
avatar  of Dannypeja
Dannypeja