Skip to content

CLI

The vaak command is installed with the package. Run vaak --help or vaak <command> --help for the built-in reference.

Terminal window
vaak transcribe clip.mp3 --model faster-whisper:small --language hi [--json]
Flag Default Meaning
FILE required Any audio or video file ffmpeg can decode
--model $VAAK_MODEL or faster-whisper:small Model spec, see model specs
--language auto-detect ISO 639-1 code: hi, en, …
--device $VAAK_DEVICE or auto auto, cpu or cuda
--json off Print the full Transcript as JSON instead of plain text

Plain output is the transcript text. JSON output has the fields text, language, segments, duration, rtf, backend, model.

Multiple files:

Terminal window
for f in *.wav; do vaak transcribe "$f" --json > "${f%.wav}.json"; done

Weights load once per process, so for many files the Python API or the server is faster.

Terminal window
vaak serve --model faster-whisper:small --host 0.0.0.0 --port 8000
Flag Default Meaning
--model $VAAK_MODEL or faster-whisper:small Model loaded at startup
--host $VAAK_HOST or 127.0.0.1 Bind address
--port $VAAK_PORT or 8000 Port
--device $VAAK_DEVICE or auto auto, cpu or cuda

Starts the FastAPI app from vaak/server/app.py. Routes are documented in the HTTP API guide. Interactive docs are at /docs once it is running.

Terminal window
vaak models

Lists every registered backend with its spec prefix, whether its runtime is importable in the current environment, and example model ids. Use it to check that an extra installed correctly.

Terminal window
vaak bench --models faster-whisper:small,indic-conformer:hi --datasets fleurs-hi,fleurs-en --limit 200
Flag Meaning
--models Comma-separated model specs
--datasets Comma-separated dataset keys: fleurs-hi, fleurs-en, cv-hi, kathbath-hi
--limit Number of utterances per dataset. Default evaluates the full test split

Evaluates every (model, dataset) pair and writes bench/results/<timestamp>.json. A model that does not support the dataset language (for example parakeet:* on fleurs-hi) is skipped. cv-hi needs a Hugging Face token in HF_TOKEN because Common Voice is gated. See the benchmarks page for the schema and methodology.

VAAK_MODEL, VAAK_DEVICE, VAAK_HOST, VAAK_PORT and VAAK_COMPUTE_TYPE set defaults for the flags above. See configuration.