Installation

vaak needs Python 3.10 or newer and ffmpeg on the PATH for audio decoding.

Base package

pip install vaak
# not on PyPI yet; until then: pip install "vaak @ git+https://github.com/mirukullacode/vaak"

or with uv:

uv pip install vaak

The base package contains the audio loader, metrics, CLI, server and the backend registry. It has no model runtime. Install one or more extras for the backends you want.

Extras

Extra	Backends enabled	Pulls in
vaak[whisper]	faster-whisper:*, hf-whisper:*, wav2vec2:*	faster-whisper, ctranslate2, transformers, torch
vaak[nemo]	indic-conformer:*, parakeet:*	nemo_toolkit[asr], torch
vaak[vosk]	vosk:*	vosk
vaak[all]	everything above	all of the above

pip install "vaak[whisper]"          # most people start here
pip install "vaak[whisper,vosk]"     # add a CPU-only fallback
pip install "vaak[all]"              # benchmarks and fine-tuning

Quote the extras so your shell does not expand the brackets.

CUDA

• faster-whisper uses CTranslate2. The PyPI wheel bundles CUDA 12 and cuDNN support; you still need an NVIDIA driver that supports CUDA 12. Set VAAK_COMPUTE_TYPE=float16 on a GPU and int8 on a CPU. See configuration.
• torch is installed by the whisper and nemo extras. pip resolves a CUDA build on Linux by default. On Windows, or if you want a specific CUDA version, install torch first from the official selector at pytorch.org and then install vaak.
• NeMo (vaak[nemo]) is large and pins its own torch range. Use a fresh virtual environment or the Docker image.
• No GPU? Everything still works on CPU. Use vosk:* or faster-whisper:tiny|base|small for acceptable speed.

Check what vaak can see:

vaak models

This lists every registered backend, whether its runtime is importable, and the device it would use.

Verify

python -c "import vaak; print(vaak.__version__)"
vaak transcribe path/to/clip.wav --model faster-whisper:tiny

The first call downloads weights from the Hugging Face Hub into the usual cache (~/.cache/huggingface). Later calls are offline.