Fine Tuning XTTS v2 for Hindi Speech with forked Coqui TTS

Channel:

NanoNomad

Subscribers:

2,970

Published on July 4, 2024 4:43:23 AM ● Video Link: https://www.youtube.com/watch?v=zts2P5X211o

Duration: 8:05

6,926 views

0

I heard a few people say that the base XTTSv2 2.0.3 model doesn't produce very good Hindi output. Here I go through preparing some Hindi speech datasets using the Common Voice version 18 and Indic TTS Database datasets. I put together a handful of metadata conversion scripts, cleanup scripts, and batch files for this.

You can copy and paste them from here:
http://nanonomad.com/2024/07/03/xttsv2-hindi-finetuning/

The fine-tuned XTTSv2 checkpoints and speaker reference clips are here:
https://huggingface.co/AOLCDROM/XTTSv2-Hi_ft/tree/main

Note: The highest step count may not necessarily be the best quality output for whatever speaker you are trying to use. Some checkpoints may repeat the end of sentences more than others.

Previous XTTS v2 video:    • Fine Tuning XTTS v2 with forked Coqui...