So last night, I realized there’s actually a TensorFlow implementation of transfer learning in DCTTS, that means I could do transfer learning for voices with less samples. The is something called real-time voice cloning which only needs an input of a 5 second audio to synthesize a text using its voice. And there’s actually someone who has combined both TTS and animation together, they called it neural voice puppetry. Well, that’s a research paper that’s relatively new and there’s no implementation publicly available yet. But it’s an interesting discovery and the result is so convincing.
And the craziest part of the day is that I LOST ALL of my data. I was going to resize the Ubuntu partition of my system. Forgetting about the graphics incompatibility in stock Ubuntu settings, the live OS that I’m running suddenly freeze during resizing the partition. Great, then I wasn’t able to boot into the system, the partition is ruined. Then something superblock recovery but totally useless. Then, I boot the live USB again with the “nouveau.modeset=0” and the crashes is gone, well, also my data. I used TestDisk, great software btw, to rebuild the partition table and recovery files from the filesystem. At first, I was amazed by the fact that my file is still here, but somehow some files are not accessible. I thought that those files are gone forever, which are the scripts that I wrote to pre-process the raw game data files. Not a big deal, but I was quite persistent, I thought maybe PyCharm has cached those files. Nope, it’s not, at least not retrievable anymore. Then I looked closer, and I realized it was a permission problem. Somehow, it was set to only readable for root. Then, reset the permissions and my scripts are back. THANK GOD!
I immediately upload them to google drive/GitHub. What a relieve. Time to proceed to experiment with the VOCA model…