docker-whisper
Self-hosted Whisper speech-to-text server with speaker diarization.
Run private speech-to-text and translation APIs with OpenAI compatibility.
This tool runs a local Whisper speech-to-text server inside Docker, exposing OpenAI-compatible transcription APIs. It suits mental health professionals who need completely private audio processing without sending sensitive client sessions to third-party clouds.
What it is
Powered by faster-whisper, it processes various audio formats, generates JSON, SRT, or VTT outputs, and supports optional speaker diarization using sherpa-onnx. You can deploy it on standard CPU hardware or accelerate inference using an NVIDIA GPU via dedicated image tags.
Who it's for
It suits practitioners and clinics requiring strict data privacy for recorded sessions who can manage containerized deployments. It does not suit non-technical users who require a ready-to-use desktop application without command-line or Docker setup.
Key features
OpenAI-Compatible APIs
Drop-in replacement endpoints for existing apps using POST /v1/audio/transcriptions.
Speaker Diarization
Identify who is speaking in each segment using optional local sherpa-onnx extensions.
SSE Streaming
Receive transcribed segments progressively as they are decoded instead of waiting for the full file.
Air-Gapped Mode
Run completely offline without internet access using pre-cached models and WHISPER_LOCAL_ONLY.
Up and running in 3 steps
Our honest verdict
What's great
- Audio stays entirely on your local server with zero third-party data transmission
- Supports all Whisper models including large-v3 and large-v3-turbo
- Provides NVIDIA GPU CUDA acceleration for faster inference
- Offers multiple subtitle and response formats including SRT, VTT, and verbose JSON
Where it falls short
- Requires technical familiarity with Docker, environment variables, and command-line interfaces
- Consumes at least 700 MB of RAM for the base model and significantly more for larger models or GPU runtimes
- Demands compatible host hardware, NVIDIA drivers, and the NVIDIA Container Toolkit for GPU acceleration
Pricing & license
Free under its license, with no paid hosted options provided by this project. The only cost is running and maintaining your own server hardware.
Open remote AI-training roles
863 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
Can I run this without an internet connection?
Yes, you can operate in offline or air-gapped mode by pre-caching models and setting WHISPER_LOCAL_ONLY=true.
What hardware do I need?
You need a Linux server with Docker, at least 700 MB of free RAM for the base model, and optionally an NVIDIA GPU with the Container Toolkit.
Does it support real-time microphone input?
This tool is designed for complete audio files; for live streaming or microphone capture, the author provides a separate docker-whisper-live tool.
