Self-Hosted Transcription

Your server, your models, your rules. Run a Whisper server on your own hardware, paste its URL into the Diction app, and your audio never leaves your network.

Good for: regulated industries where audio cannot go to the cloud, people who already run servers at home, anyone who wants a specific model for a specific language or domain, and folks who refuse to hand their voice to someone else's infrastructure.

How it works

Diction speaks the OpenAI transcription API directly. If your server accepts POST /v1/audio/transcriptions, Diction can talk to it. That's the whole contract.

The easiest way is our open-source stack: a speech engine plus the Diction gateway. Two commands on any machine that runs Docker, no GPU needed:

git clone --depth 1 https://github.com/DictionLabs/Diction && cd Diction
docker compose --profile parakeet up -d

Then scan the gateway's QR code in the app and you're dictating. The gateway streams your audio live while you're still talking, so by the time you tap stop, the transcript is already coming back.

The default engine covers 25 European languages and runs fast on a plain CPU. Need one of the other 74? Switch to a Whisper profile, which covers 99 languages. The setup guide has both.

Full walkthrough: Self-Hosting Setup Guide. Already running your own Whisper server? Use Your Own Model.

What you get

• Audio stays on your network. From your phone to your server and back. No third party in the middle.
• Free, unlimited, no restrictions. No Diction subscription needed. No word limits. No daily caps. No trial that expires.
• Works with any Whisper-compatible server. The app speaks the OpenAI transcription API directly. Use our default stack, use someone else's, roll your own.
• The same Writing Tools as Diction Cloud. Point the gateway at any OpenAI-compatible LLM, a local one included, and cleanup, voice editing and suggestions run on it. 7B or larger recommended; smaller models often don't follow instructions.
• Streaming. The gateway streams audio as you speak. Longer dictations are noticeably faster.
• On-device fallback. If your server is unreachable, Diction automatically retries using a local model on your iPhone. Your dictation is never lost to a network issue.
• Run it anywhere. Home server, NAS, Raspberry Pi 4 or 5, cloud VM, behind a reverse proxy, over a VPN. If Docker runs there, Diction connects to it.
• Pick your own model. Run whatever speech model fits your use case. Your language, your accuracy requirements, your hardware.
• Open-source gateway. The gateway infrastructure is fully open source on GitHub. Inspect it, modify it, contribute to it.

Best for

• You already run Docker at home and want transcription on your own hardware
• You work in a regulated industry where audio cannot leave your network
• You want a specific model for a specific language or domain
• You refuse to send voice data to someone else's cloud
• You already have a Whisper server running and just want an iOS keyboard that talks to it