AgnesTachyon So-vits-svc 4.1 Model

A so-vits-svc 4.1 model of AgnesTachyon in Uma Musume: Pretty Derby.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. --> This is a so-vits-svc 4.1 model of AgnesTachyon in Uma Musume: Pretty Derby.

• Developed by: svc-develop-team
• Trained by: 70295
• Model type: Audio to Audio
• License: CC BY-NC 4.0

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

• Clone the so-vits-svc repository and install all dependencies.
• Create a new folder named "models" and place the "AgnesTachyon" folder inside it.
• Navigate to the directory of "so-vits-svc" and execute the following command by replacing "xxx.wav" with the name of your source audio file and "x" with the desired key to raise/lower.

python inference_main.py -m "models/AgnesTachyon/AgnesTachyon.pth" -c "models/AgnesTachyon/config.json" -n "xxx.wav" -t x -s "AgnesTachyon"

Shallow diffusion model, cluster model and feature index model is also provided. Check the README.md file of the so-vits-svc project for more information.

Training Details

Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. --> All of the training data is extracted from the Windows client of Uma Musume: Pretty Derby using the umamusume-voice-text-extractor.
The copyright of the training dataset belongs to Cygames.
Only the voice is used, the live music soundtrack is not included in the training dataset.

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

Training Environment Preparation

• Download the base models mentioned in the README.md file of the so-vits-svc project.
  You should download checkpoint_best_legacy_500.pt , D_0.pth, G_0.pth(for sovits model), model_0.pt(for shallow diffusion) , rmvpe.pt(for the f0 predictor RMVPE), model(for NSF_hifigan).
• Place checkpoint_best_legacy_500.pt, rmvpe.pt in .\pretrain, place model and its config.json in .\pretrain\nsf_hifigan, place D_0.pth, G_0.pth in .\logs\44k, place model_0.pt in .\logs\44k\diffusion .
  Credits: The D_0.pth and G_0.pth provided above is from OOPPEENN.

Preprocessing

• Delete all WAV files smaller than 400KB, and copy them to .\dataset_raw\AgnesTachyon
• Navigate to the directory of "so-vits-svc" and execute python resample.py --skip_loudnorm .
• Execute python preprocess_flist_config.py --speech_encoder vec768l12 --vol_aug .
• Edit the parameters in config.json and diffusion.yaml.
• Execute python preprocess_hubert_f0.py --f0_predictor rmvpe --use_diff

Training

• Execute python train.py -c configs/config.json -m 44k .

[Optional]

• Execute python train_diff.py -c configs/diffusion.yaml to train the shallow diffusion model.
• Execute python cluster/train_cluster.py --gpu to train the cluster model.
• Execute python train_index.py -c configs/config.json to train the feature index model.

Training Hyperparameters

Please check config.json and diffusion.yaml for training hyperparameters

Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: RTX 3090
• Hours used: 41.6
• Provider: Myself
• Compute Region: Mainland China
• Carbon Emitted: ~16.02kg CO2