Installation¶

Prerequisites¶

Python 3.12 or newer¶

python --version

ffmpeg (required)¶

Extracts audio from video and prepares files for NeMo.

Windows:

winget install Gyan.FFmpeg

Verify:

ffmpeg -version
ffprobe -version

curl (for model downloads on Windows)¶

Included in Windows 10+. Scripts use curl.exe explicitly.

Local installation (recommended on Windows)¶

From the repository root:

python scripts\install_local.py

The script:

Creates .venv/ if missing
Installs PyTorch (CPU)
Installs Sbobinator with pip install -r requirements/local.txt (NeMo, FastAPI, LLM summary)

Optional dependencies (`pyproject.toml`)¶

Extra	Packages	Use
`asr`	torch, nemo_toolkit	Transcription only
`ui`	fastapi, uvicorn, jinja2, cloud LLM clients	Web UI + API summary
`summarize`	openai, anthropic, google-genai, llama-cpp-python	LLM summary modules
`local`	all of the above	Full installation
`dev`	ruff, mkdocs	Lint and docs

Manual installation¶

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux:   source .venv/bin/activate
pip install --upgrade pip
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -e ".[local]"

Verify installation¶

sbobina info

Should show version, default model, and detected device.

What NOT to use¶

No PowerShell scripts

.ps1 scripts have been removed (antivirus false positives). Use only:

python scripts/*.py
start.bat
sbobina CLI