Finally, I made another character’s voice (Yennefer) which has only around 1k samples. I’ve heard of that video mentioning about transfer learning and I was so eager to know about his code since the result was quite good, but the guy didn’t provide any code. Then, I’ve found an implementation of a transfer learning like the guy did. Unfortunately, TensorFlow magically doesn’t work well on my computer (it crashes for some reason, maybe my CPU is too old). Then, I’ve looked at some forks of the original TensorFlow implementation, one of the repo mentioned that you can actually train on a pretrained model, it not only works across gender, but different accents as adapted well. So here it is, no fancy and complicated technique, just keep on training. It already sounds so natural even without training the SSRN! This is the result of only 10K steps (less than an hour), it changes the accent! This is indeed magic! So natural! I think I can do other characters’ voices now!
Yennefer’s British voice:
Samples:
Pretrained:
Transferred:
The birch canoe slid on the smooth planks.
Glue the sheet to the dark blue background.
Rice is often served in round bowls.
The box was thrown beside the parked truck.
Help the woman get back to her feet.
A pot of tea helps to pass the evening.
The salt breeze came across from the sea.
The girl at the booth sold fifty bonds.
What a surprise! I’m so excited for the GPT-2 part and building a web app!