Skip to content

Latest commit

 

History

History
72 lines (48 loc) · 3.12 KB

README.md

File metadata and controls

72 lines (48 loc) · 3.12 KB
Raw

xtts-finetune-webui

This is my customized version of xtts-finetune-webui.

I added support for custom speaker names, plentiful debugging in the console, upgrade gradio lib to bypass a problem, and most critically, buffering on the audio files to minimize the prematurely truncated sentences due to crummy timestamps from whisper.

TODO

  • Add the ability to use via console

Key features:

Data processing

  1. Updated faster-whisper to 0.10.0 with the ability to select a larger-v3 model.
  2. Changed output folder to output folder inside the main folder.
  3. If there is already a dataset in the output folder and you want to add new data, you can do so by simply adding new audio, what was there will not be processed again and the new data will be automatically added
  4. Turn on VAD filter
  5. After the dataset is created, a file is created that specifies the language of the dataset. This file is read before training so that the language always matches. It is convenient when you restart the interface

Fine-tuning XTTS Encoder

  1. Added the ability to select the base model for XTTS, as well as when you re-training does not need to download the model again.
  2. Added ability to select custom model as base model during training, which will allow finetune already finetune model.
  3. Added possibility to get optimized version of the model for 1 click ( step 2.5, put optimized version in output folder).
  4. You can choose whether to delete training folders after you have optimized the model
  5. When you optimize the model, the example reference audio is moved to the output folder
  6. Checking for correctness of the specified language and dataset language

Inference

  1. Added possibility to customize infer settings during model checking.

Other

  1. If you accidentally restart the interface during one of the steps, you can load data to additional buttons
  2. Removed the display of logs as it was causing problems when restarted
  3. The finished result is copied to the ready folder, these are fully finished files, you can move them anywhere and use them as a standard model
  4. Added support for finetune Japanese

Changes in webui

1 - Data processing

image

2 - Fine-tuning XTTS Encoder

image

3 - Inference

image

Install

  1. Make sure you have Cuda installed
  2. git clone https://github.com/daswer123/xtts-finetune-webui
  3. cd xtts-finetune-webui
  4. pip install torch==2.1.1+cu118 torchaudio==2.1.1+cu118 --index-url https://download.pytorch.org/whl/cu118
  5. pip install -r requirements.txt

If you're using Windows

  1. First start install.bat
  2. To start the server start start.bat
  3. Go to the local address 127.0.0.1:5003

On Linux

  1. Run bash install.sh
  2. To start the server start start.sh
  3. Go to the local address 127.0.0.1:5003 ~