Whisper Server (Docker)

A Whisper transcription service using faster-whisper that exposes an OpenAI-compatible API. Runs on any platform via Docker (CPU-only, no GPU required).

This is the Docker/Railway-deployable counterpart to whisper-server-apple-silicon.

Quick Start

Docker

docker build -t whisper-server .
docker run -p 8080:8080 whisper-server

The model (large-v3-turbo) is downloaded on first startup and cached inside the container. To persist the cache across restarts:

docker run -p 8080:8080 -v whisper-cache:/root/.cache whisper-server

Railway

Deploy directly from this repo. Railway will use the Dockerfile automatically. No additional configuration needed — the server listens on port 8080 by default.

Configuration

Environment variables:

Variable	Default	Description
WHISPER_SERVER_HOST	0.0.0.0	Host to bind to
WHISPER_SERVER_PORT	8080	Port to listen on

API Usage

Transcribe Audio

curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F "file=@audio.mp3"

Response:

{"text": "Transcribed text here..."}

Health Check

curl http://localhost:8080/health

Model

Uses large-v3-turbo via faster-whisper (CTranslate2 backend). The model is downloaded from Hugging Face on first use and cached in ~/.cache/huggingface.

License

MIT