MANUAL.md

Herta Voice Changer

Introduction

This AI model is based on SoftVC VITS Singing Voice Conversion. Refer to this Github Repository from the 4.0 branch. This model was inspired by Herta from Honkai Star Rail. This model can be used to convert the original voice from an audio file into this character's voice.

How to Prepare Audio Files

Your audio files should be shorter than 10 seconds, have no BGM, and have a sampling rate of 44100 Hz.

1. Create a new folder inside the dataset_raw folder (This folder name will be your SpeakerID).
2. Put your audio files into the folder you created above.

Note:

1. Your audio files should be in .wav format.
2. If your audio files are longer than 10 seconds, I suggest you trim them down using your desired software or audio slicer GUI.
3. If your audio files have BGM, please remove it using a program such as Ultimate Vocal Remover. The 3_HP-Vocal-UVR.pth or UVR-MDX-NET Main is recommended.
4. If your audio files have a sampling rate different from 44100 Hz, I suggest you resample them using Audacity or by running python resample.py in your CMD.

How to Build Locally

1. Clone the repository from the 4.0 branch: git clone https://github.com/svc-develop-team/so-vits-svc.git
2. Put your prepared audio into the dataset_raw folder.
3. Open your Command Line and install the so-vits-svc library: %pip install -U so-vits-svc-fork
4. Navigate to your project directory using the Command Line.
5. Run svc pre-resample in your prompt.
6. After completing the step above, run svc pre-config.
7. After completing the step above, run svc pre-hubert. (This step may take a while.).
8. After completing the step above, run svc train -t. (This step will take a while based on your GPU and the number of epochs you want.).

How to Change Epoch Value Locally

The meaning of epoch is the number of training iterations for your model. Example: if you set the epoch value to 10000, your model will take 10000 steps to finish (default epoch value is 10000). To change your epoch value:

1. Go to your project folder.
2. Find the folder named config.
3. Inside that folder, you should see config.json.
4. In config.json, there should be a section that looks like this:

    "train": {
        "log_interval": 200,
        "eval_interval": 800,
        "seed": 1234,
        "epochs": <PUT YOUR VALUE HERE>,
        "learning_rate": 0.0001,
        "betas": [0.8, 0.99]
    }

This can be done after svc pre-config has already finished.

How to inferance in local.

To perform inference locally, navigate to the project directory, create a Python file, and copy the following lines of code:

your_audio_file = 'your_audio.wav'

audio, sr = librosa.load(your_audio_file, sr = 16000, mono = True)
raw_path = io.BytesIO()
soundfile.write(raw_path, audio, 16000, format = 'wav')
raw_path.seek(0)

model = Svc('logs/44k/your_model.pth', 'logs/44k/config.json')

out_audio, out_sr = model.infer('<YOUR SPEAKER ID>', 0, raw_path, auto_predict_f0 = True)
soundfile.write('out_audio.wav', out_audio.cpu().numpy(), 44100)

The output file will be in the same directory as your input audio file with the name your_audio_out.wav

How to Build in Google Colab

Refer to My Google Colab or the Official Google Colab for the steps.

Google Drive Setup

1. Create an empty folder (this will be your project folder).
2. Inside the project folder, create a folder named dataset_raw.
3. Create another folder inside dataset_raw (this folder name will be your SpeakerID).
4. Upload your prepared audio files into the folder created in the previous step.

Google Colab Setup

1. Mount your Google Drive:

   from google.colab import drive  
   drive.mount('/content/drive')  
   

2. Install dependencies:

   !python -m pip install -U pip setuptools wheel  
    %pip install -U ipython  
    %pip install -U torch torchaudio --index-url https://download.pytorch.org/whl/cu118  
   

3. Install so-vits-svc library: %pip install -U so-vits-svc-fork

4. Resample your audio files: !svc pre-resample

5. Pre-config: !svc pre-config

6. Pre-hubert (this step may take a while): !svc pre-hubert

7. Train your model (this step will take a while based on your Google Colab GPU and the number of epochs you want): !svc train -t

How to Change Epoch Value in Google Colab

The term "epoch" refers to the number of times you want to train your model. For example, if you set the epoch value to 10,000, your model will take 10,000 steps to complete (the default epoch value is 10,000).

To change the epoch value:

1. Go to your project folder.
2. Find the folder named config.
3. Inside that folder, you should see config.json.
4. In config.json, there should be a section that looks like this:

    "train": {
        "log_interval": 200,
        "eval_interval": 800,
        "seed": 1234,
        "epochs": <PUT YOUR VALUE HERE>,
        "learning_rate": 0.0001,
        "betas": [0.8, 0.99]
    }

This can be done after svc pre-config has already finished.

How to Perform Inference in Google Colab

After training your model, you can use it to convert any original voice to your model voice by running the following command:

!svc infer drive/MyDrive/your_model_name/your_audio_file.wav --model-path drive/MyDrive/your_model_name/logs/44k/your_model.pth --config-path drive/MyDrive/your_model_name/logs/44k/your_config.json  

The output file will be named your_audio_file.out.wav

Note:

1. Your Google Drive must have at least 5 GB of free space. If you don't have enough space, consider registering a new Google account.
2. Google Colab's Free Subscription is sufficient, but using the Pro version is recommended.
3. Set your Google Colab Hardware Accelerator to GPU.

Credits

1. zomehwh/sovits-models from Hugging Face Space
2. svc-develop-team/so-vits-svc from GitHub repository
3. voicepaw/so-vits-svc-fork from GitHub repository