Synthesizer doesn't work in Python 3.8+

CorentinJ / Real-Time-Voice-Cloning

Clone a voice in 5 seconds to generate arbitrary speech in real-time

Other

52.54k stars 8.79k forks source link

Synthesizer doesn't work in Python 3.8+ #634

Closed ghost closed 3 years ago

ghost commented 3 years ago

The synthesizer uses Tensorflow 1.x, which is not supported in Python 3.8 or higher. We need help to upgrade the synthesizer model.

In #370, we tried upgrading the code to Tensorflow 2.x. However, the differences between TF 1.x and 2.x are significant. Even switching to an existing TF 2.x Tacotron model would require an overhaul of the training and inference scripts.
The preferred approach is to replace it with a Pytorch-based synthesizer. This is something we discussed in #447. I submitted a PR #472, which was not accepted due to bad voice quality. The heavy lifting is done but more work is needed to improve the model.

Please respond to this issue or submit a pull request to help out.

vinamramunot-tech commented 3 years ago

I am also facing the same issue. I tried to see if there was packages in tensorflow_addon however couldn't find it. Is there an update to this? @blue-fish

ghost commented 3 years ago

The information above is still current. My fork works with Python 3.8, but it's not the long-term solution to this problem. Someone else needs to step up and contribute a better model.

vinamramunot-tech commented 3 years ago

@blue-fish Although I am not knowledgeable but I can give it a try. Is there something that we can start on?

ghost commented 3 years ago

@vinamramunot-tech Actually I just found something that should improve my pytorch model. I'll report back in a few days if it worked. https://github.com/CorentinJ/Real-Time-Voice-Cloning/pull/472#issuecomment-765875497

vinamramunot-tech commented 3 years ago

@blue-fish thank you!! Let me know if I can be of any help.

ghost commented 3 years ago

@vinamramunot-tech It worked, and I expect to have a model for release within a week.

If you are available to help, it would be nice to add mixed precision training support. That will allow us to train better models on consumer GPUs with limited memory. If you have any interest in helping, leave a message in #487 and I'll let you know how you can get started.

ghost commented 3 years ago

Tentatively closing this since this issue doesn't require attention at this time.