vits-finetuning

Why we included this project

Building a custom text-to-speech voice usually means training a model from scratch, which is expensive. This project skips that by fine-tuning VITS from a pre-trained checkpoint, and it documents the whole process: preparing a Japanese audio-text dataset, resampling clips to the required format, preprocessing, and multi-speaker training. The README is honest about data needs, noting that roughly fifty audio-text pairs can already give decent results, which makes it practical for hobbyists and small teams chasing a character or personal voice. A Google Colab notebook covers both training and inference, so you don't need your own GPU. The main limitation is that only Japanese datasets are supported, so it only.