Start with mechanics, then connect models

Install

Use Python 3.10 or newer. The lightweight package does not install CUDA models. Run commands from the repository root unless a guide says otherwise.

python3 -m venv .venv
source .venv/bin/activate
pip install -e '.[dev,evaluation]'
aegisvox --help

The base dependency is PyYAML, plus the Python audioop compatibility package on Python 3.13+. The evaluation extra adds NumPy, pandas, SciPy, and statsmodels. The API extra adds FastAPI, Uvicorn, and HTTPX. Install each model's dependencies in a separate environment to avoid incompatible Torch, audio, and codec packages.

Run the offline example

aegisvox demo --output runs/demo-001
cat runs/demo-001/result/reconstruction.json
pytest -q

Use a new directory for each run. The command refuses to overwrite existing output. Listen only to verify a signal exists: these are tones, not generated speech. sample_equivalent: true means the inserted PCM samples equal the rendered file's samples. It says nothing about the words, identity, or clinical meaning.

Supply actual speech

Prepare a PCM WAV conversation after voice conversion. Create a manifest matching examples/complete_turns.json. Times refer to that converted waveform, not to a pre-conversion waveform whose length may differ. Each edited entry contains the whole turn's source and revised transcript. The complete converted turn is cropped as the renderer reference. No user voice sample is bundled.

aegisvox render-turns /secure/work/turns.json \
  --renderer examples/f5_renderer.json --output /secure/work/render-001

Configure the renderer's executable to its dedicated model environment. Real-model execution is not covered by the offline test; see backend provisioning. Rendering and evaluation are separate operations. Do not mistake assembled for a privacy, clinical, or production-readiness verdict.