Skip to content

Run it yourself

Minimum Tested on
GPU NVIDIA, 6 GB VRAM (CUDA) RTX 3050 Laptop 6 GB
RAM 8 GB 16 GB
Disk ~10 GB (environment + model weights)
Software Python 3.10, FFmpeg on PATH, Node.js 22+ Windows 11
  1. Clone the repository

    Terminal window
    git clone https://github.com/Aayush-Dubey123/audiopipeline.git
    cd audiopipeline
  2. Create the Python environment (one environment, inside backend/)

    Terminal window
    python -m venv backend\.venv
    backend\.venv\Scripts\Activate.ps1
  3. Install PyTorch with CUDA, then the project

    Terminal window
    pip install torch==2.6.0+cu124 torchaudio==2.6.0+cu124 --extra-index-url https://download.pytorch.org/whl/cu124
    pip install -r requirements.txt
  4. Configure by copying backend/.env.example to backend/.env. The defaults work; the voice-tuning settings are explained in Tuning by measurement.

  5. Start everything

    Terminal window
    .\run.ps1

    This opens the backend on http://127.0.0.1:8000 (API docs at /docs) and the studio on http://localhost:3000. The first generation downloads the model weights (a few GB) from Hugging Face.

Terminal window
backend\.venv\Scripts\python.exe -m pytest backend/tests -q

The suite uses a throwaway database and storage folder, so it leaves nothing behind.

All three settings live in backend/.env:

CHATTERBOX_EXAGGERATION=0.5 # expressiveness; keep ≤ 0.6 to avoid babbling
CHATTERBOX_CFG_WEIGHT=0.4 # lower follows the reference's pace and accent more closely
CHATTERBOX_TEMPERATURE=0.75 # higher = more variation between chunks