Skip to content

    Quickstart

    Quickstart

    Section titled “Quickstart”

    Get from zero to a working transcription in a few minutes.

    Prerequisites

    Section titled “Prerequisites”
    • Docker with the NVIDIA Container Toolkit for GPU inference on an x86_64 NVIDIA GPU (RTX 3000 Ampere → RTX 5000 Blackwell). A CPU-only run works too, with reduced throughput.
    • Outbound access to ghcr.io to pull the image. The default model is baked into the image, so no model download is needed at container start.
    • An audio file to transcribe (wav, mp3, m4a, flac, ogg, webm, …).

    1. Run the container

    Section titled “1. Run the container”
    Terminal window
    docker run -d --name whisper --gpus all -p 8000:8000 \
      ghcr.io/forgeguard-ai/faster-whisper-server:latest

    This publishes the service on port 8000 and bakes the large-v3 model. On Jetson Orin, use the Jetson image and NVIDIA runtime instead:

    Terminal window
    docker run -d --name whisper --runtime nvidia -p 8000:8000 \
      ghcr.io/forgeguard-ai/faster-whisper-server-jetson:latest

    The Jetson image bakes the small model — see Hardware profiles.

    2. Verify health and readiness

    Section titled “2. Verify health and readiness”

    The server accepts connections immediately and warms the model in the background. Two endpoints report distinct states:

    Terminal window
    curl http://localhost:8000/health   # 200 right away: {"status":"warming",...} then "healthy"
    curl http://localhost:8000/ready    # 503 while warming, 200 once it can transcribe

    Wait for /ready to return 200 {"status":"ready"} before sending inference requests. The full contract is documented in Health and readiness.

    3. Make your first transcription request

    Section titled “3. Make your first transcription request”
    Terminal window
    curl -X POST http://localhost:8000/v1/audio/transcriptions \
      -F 'file=@audio.mp3' \
      -F 'response_format=text'

    A plain-text transcript is returned. Try response_format=json for {"text": "..."}, or verbose_json for segments and optional word timestamps — see Transcription and translation.

    4. Use the OpenAI SDK

    Section titled “4. Use the OpenAI SDK”

    Any OpenAI audio client works by pointing base_url at the server:

    from openai import OpenAI
    

    client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
    with open("audio.mp3", "rb") as f:
        transcript = client.audio.transcriptions.create(model="whisper-1", file=f)
    print(transcript.text)

    The model field is accepted for compatibility but ignored; the loaded Whisper model is used. When authentication is disabled, any api_key value is accepted.

    5. Open the web console

    Section titled “5. Open the web console”

    Browse to http://localhost:8000/web to upload or record audio, pick a language and response format, watch GPU telemetry, and switch models. Interactive API docs are at http://localhost:8000/docs.

    Next steps

    Section titled “Next steps”