Fine-tuned MiMo Audio to accept text/emotion captions (e.g. "intense fury, rage, hate") as input, trained a LoRA for 1k steps on LAION's voice acting dataset.
Thanks to HF for the GPUs to train ๐ค
Hi, do you see a limit in the number of voices I have 416 and it fails to load all of them. (scroll menu limit?)
I'm not sure if there's a set limit for the dropdown, but with that many voices, it might make sense to not use the dropdown but instead have a textbox to specify the path to the reference speaker.
I don't think that's supported by the model, but you could fine-tune it or clone a voice with emotions. (I am not the author of the model itself, just of the web demo)
Hi,
You can upload a WAV file to the voices folder. Then, in the app.py file, add the filename of the voice (without .wav) to the voicelist list. It should show up in the Gradio demo.
Hi,
I added:
import nltk
nltk.download('punkt_tab')
and it seems to resolve the issue for me. Have you changed any code from the original Space?
Thanks!
Hi,
Sorry about the issues! Please try adding:
nltk.download('punkt_tab')
below the nltk.download() line โ let me know if it works!
Training itself would be pretty easy, but the main issue would be data. AFAIK there's not much data out there for other TTS models. I synthetically generated the StyleTTS 2 dataset as it's quite efficient but other models would require much more compute.
Congratulations!
