Voice Training on GPU

To achieve high-quality, expressive voice cloning with custom prosody and emotion, you can fine-tune GPT-SoVITS on your own voice samples.

Reference Dataset Preparation

1. Record 1 to 3 minutes of clean, dry voice audio using an external microphone.
2. Segment the recording into 5 to 10 second WAV clips.
3. Generate exact transcripts for each clip.

1-Click Google Colab Training

Rather than running heavy PyTorch CUDA training on your local laptop, use the bundled Google Colab notebook:

Colab Training Steps:

1. Open notebooks/01_voice_clone_sovits_training.ipynb in Colab.
2. Select Runtime $\rightarrow$ Change runtime type $\rightarrow$ T4 GPU.
3. Upload your audio clips and run the automated pipeline (Hubert feature extraction, Semantic token extraction, and SoVITS fine-tuning).
4. Download the generated weights (my_voice.pth and my_voice.ckpt).

Installing Trained Weights Locally

Place your trained weights in backend/voice_samples/ and update your .env:

REF_AUDIO_PATH=backend/voice_samples/my_voice_ref.wav
REF_TEXT="Hello, this is my trained custom voice speaking."

Restart the voice_agent container:

docker compose -f docker-compose.infra.yml -f docker-compose.prod.yml restart voice_agent