Proof of concept testing Tortoise TTS with a new Latin-1/ISO8559-1 character-only BPE tokenizer.
Tutorial video up now: https://youtu.be/o7QRbMvFPzs
Model was fine-tuned with the new tokenizer with an English-language dataset (approx 13k samples), then that same model was fine tuned with a French dataset (approx 9k samples).
I only understand a little French, so it is difficult for me to qualitatively assess the output.
This isn't anywhere near perfect. It is very undertrained. Just a proof of concept using the same tokenizer and pretrained base to train other languages with smaller datasets.
There are some errors and misspoken phrases here. I've left them in, because they're funny.
And yes, the digits need to be transcribed, or the normalization needs to be disabled. All the digits here are being read in English