OpenAI Whisper Docker Image (GPU Accelerated)

Disclaimer: This is an unofficial, community-maintained Docker wrapper around OpenAI Whisper. It is not affiliated with or endorsed by OpenAI. This project is maintained as a personal hobby — use it as-is, no support is guaranteed.

This Docker image provides a convenient environment for running OpenAI Whisper, a powerful automatic speech recognition (ASR) system. Based on Ubuntu 24.04 with Python 3.12 and all necessary dependencies included.

Prerequisites

• Docker
• NVIDIA Container Toolkit (for GPU mode only)

Quick Start

1. Place your audio files in the audio-files/ directory
2. Run the transcription (see examples below)
3. Find the output in the same audio-files/ directory

Note: the first run takes a while: Whisper downloads the selected model (~1.6 GB for turbo, ~3 GB for large-v3) into the models/ directory before transcribing. Subsequent runs reuse the cached model and start right away.

Usage

Prebuilt Image

A prebuilt image is published to GitHub Container Registry on every push to main and rebuilt monthly:

docker pull ghcr.io/manzolo/openai-whisper-docker:latest

If you only run on CPU, pull the much smaller CPU-only variant (~3.5 GB instead of ~13 GB, PyTorch without CUDA):

docker pull ghcr.io/manzolo/openai-whisper-docker:cpu

Date-stamped tags (e.g. :20260705, :cpu-20260705) are also available if you want to pin a specific build.

Docker Compose (Recommended)

Build:

docker compose --profile gpu build

GPU mode:

docker compose run --rm whisper-gpu whisper audio-file.mp3 --device cuda --model turbo --language Italian --output_dir /app --output_format txt

CPU mode:

docker compose run --rm whisper-cpu whisper audio-file.mp3 --model turbo --language Italian --output_dir /app --output_format txt

Docker Direct

Build:

docker build -t openai-whisper .

GPU mode:

docker run --gpus all -it -v ${PWD}/models:/root/.cache/whisper -v ${PWD}/audio-files:/app openai-whisper whisper audio-file.mp3 --device cuda --model turbo --language Italian --output_dir /app --output_format txt

CPU mode:

docker run -it -v ${PWD}/models:/root/.cache/whisper -v ${PWD}/audio-files:/app openai-whisper whisper audio-file.mp3 --model turbo --language Italian --output_dir /app --output_format txt

Volume Mounts

Local Path	Container Path	Description
./models	/root/.cache/whisper	Cached Whisper models (persisted between runs)
./audio-files	/app	Input audio files and output transcriptions

Models

Model	VRAM Required	Description
large-v3	10-15 GB	Most accurate, recommended for powerful GPUs
turbo	~8 GB	Memory-efficient, near-comparable accuracy

Useful Commands

Check GPU information:

docker run --gpus all -it openai-whisper nvidia-smi

License

This project is licensed under the MIT License. The image bundles third-party software under its own licenses: OpenAI Whisper (MIT), PyTorch (BSD-3), ffmpeg (from Ubuntu packages, LGPL/GPL), and Ubuntu.

References

• OpenAI Whisper GitHub Repository
• NVIDIA Container Toolkit Installation Guide
• Docker Official Website

---

🧠 Local AI Lab

This project is part of manzolo's Local AI Lab — a family of self-hosted AI projects (LLM, voice, vision & documents) that share the same conventions and can be wired together through the shared local-ai-net Docker network.

Explore the whole family: topic:local-ai