Aw, well thank you ! I saw your other ask offering paying me— that is not necessary at all. My engineer model has the following data: His original voice lines, the Meet Your Match voice lines, the Paladins crossover voice lines. All together, I think that comes out to around 20 minutes. I then added an extra 5 minutes of generated voice lines from Uberduck TTS (just reading random wikipedia pages). I didn't add any more data than that, because I found that more advanced models tend to remove the accent of the original voice. You can make as much artificial data as you want and it will become smoother, but for me I wanted to prioritize the character and I didn't want it to be a perfect recreation of the voice actor because it's a jinriki. My model is still pretty glitchy honestly, but it helps to also pay attention to the Tensorboard. I don't fully understand how to read it, but I have attached an image trying to explain the val/mel chart, which is what I use to judge the quality. Throughout training, your model will go up and down in quality, and it's possible to also overtrain it and actually make it worse. You can usually tell that you've reached peak quality when you hit a peak that is suddenly really high, and it doesn't get up to that height again even a long time after that. My engineer model stopped somewhere around 69-72k i think. With a model that's so few minutes, it will probably stop around that time too. I hope that helps