Configuration

vaak has no config file. Defaults come from environment variables; command-line flags override them.

Variable	Default	Used by	Meaning
VAAK_MODEL	faster-whisper:small	transcribe, serve, Space	Default model spec
VAAK_DEVICE	auto	all backends	auto, cpu or cuda. auto uses CUDA when torch or CTranslate2 can see a GPU
VAAK_HOST	127.0.0.1	serve	Bind address. Use 0.0.0.0 in containers
VAAK_PORT	8000	serve	TCP port
VAAK_COMPUTE_TYPE	float16 on CUDA, int8 on CPU	faster-whisper	CTranslate2 compute type: float16, int8_float16, int8, float32

Examples

export VAAK_MODEL=indic-conformer:hi
export VAAK_DEVICE=cuda
vaak serve --host 0.0.0.0

VAAK_MODEL=vosk:vosk-model-small-hi-0.22 vaak transcribe clip.wav

Other variables vaak respects

These belong to upstream libraries but matter in practice:

Variable	Meaning
HF_HOME	Where Hugging Face caches weights. Default ~/.cache/huggingface
HF_TOKEN	Needed for gated datasets such as Common Voice (cv-hi)
HF_HUB_OFFLINE	Set to 1 to refuse network access once weights are cached
CUDA_VISIBLE_DEVICES	Pin a GPU when the machine has several
OMP_NUM_THREADS	CPU thread count for CTranslate2 and torch

Precedence

1. Command-line flag (--model, --device, --host, --port)
2. Environment variable
3. Built-in default

The HTTP model form field overrides VAAK_MODEL per request and caches the extra backend in the server process.