Wyoming Faster Whisper (CUDA)

Wyoming protocol server for faster-whisper speech-to-text, running on NVIDIA GPU via CUDA.

Forked from rhasspy/wyoming-faster-whisper. For CPU or Home Assistant add-on use, see the upstream repo.

Prerequisites

• NVIDIA GPU with CUDA 12.6 compatible driver
• nvidia-container-toolkit installed on the host

Quick Start

docker compose up

Edit docker-compose.yml to set your data volume path and preferred model before running.

Manual Docker Run

docker build -t wyoming-faster-whisper:cuda .

docker run -it --gpus all \
    -p 10300:10300 \
    -v /path/to/local/data:/data \
    wyoming-faster-whisper:cuda \
    --device cuda \
    --model tiny-int8 \
    --language en

Configuration

Key options passed after the image name:

Option	Description
--device cuda	Use GPU for inference (required)
--model	Model name or HuggingFace ID (e.g. tiny-int8, small-int8, Systran/faster-distil-whisper-small.en)
--language	Language code (e.g. en), or omit for auto-detect
--beam-size	Beam size for decoding (default: 1)
--data-dir	Directory where models are stored (default: /data)

Models are downloaded to /data on first run — mount a persistent volume there.

Port

The wyoming server listens on TCP port 10300.