Quickstart

1. Install

pip install "vaak[whisper]"

ffmpeg must be on your PATH. See installation for CUDA and other extras.

2. Transcribe a file

vaak transcribe clip.mp3 --language hi

Prints the transcript to stdout. The default model is faster-whisper:small. Add --json to get the full Transcript object including segments and the real-time factor:

vaak transcribe clip.mp3 --model faster-whisper:small --language hi --json

Leave out --language to auto-detect.

3. Use it from Python

from vaak import get_backend, load_audio

audio = load_audio("clip.wav")                    # float32 numpy, 16 kHz mono
backend = get_backend("faster-whisper:small")      # lazy: weights load on first use
result = backend.transcribe(audio, language="hi")  # language=None -> auto-detect
print(result.text)
print(result.language, result.rtf)

result.segments is a list of Segment(start, end, text). See the Python API guide.

4. Start a server

vaak serve

By default this loads faster-whisper:small and listens on http://127.0.0.1:8000. Try it:

curl -F file=@clip.mp3 -F language=hi http://localhost:8000/v1/transcribe

Response:

{
  "text": "...",
  "language": "hi",
  "segments": [{"start": 0.0, "end": 2.4, "text": "..."}],
  "duration": 2.4,
  "rtf": 0.08,
  "backend": "faster-whisper",
  "model": "small"
}

The server also exposes /v1/audio/transcriptions, which accepts the same request shape as OpenAI’s endpoint, and a WebSocket at /v1/stream. See the HTTP API guide.

5. Try another backend

pip install "vaak[nemo]"
vaak transcribe clip.mp3 --model indic-conformer:hi

vaak models lists everything that is installed. Spec strings are documented in model specs.

Next

• Live demo in the browser, or pointed at the server you just started.
• Benchmarks to pick a model for your language and hardware.
• Docker to run the server without a Python environment.