OpenWispr documentation

What OpenWispr is

OpenWispr is a dictation app for Android and macOS. You trigger it, speak, and the text is written into whatever field you were already in. Speech recognition and text cleanup both run on your device by default: on a fresh install there is no account, no server, and nothing to sign in to, and after the one-time model download the app does not need a network connection to transcribe anything.

It is MIT-licensed, and the whole of it is on GitHub. Everything on these pages was written by reading that source. Where the app's behaviour is surprising, these pages say so rather than smoothing it over, and where the code does not settle a question — model language coverage is the clearest example — they say that too instead of guessing.

What happens when you dictate

The same pipeline runs on both platforms. Each stage is documented in more detail on the page linked beside it.

• Trigger. On Android, the floating bubble: tap it for a long-form take, or hold it and speak and release to send. On macOS, hold the fn/globe key, or use a global shortcut you record yourself. See Android setup and macOS setup.
• Record. 16 kHz mono PCM, with a Silero voice-activity model watching the stream so a tap-to-start recording can end itself when you stop talking.
• Transcribe, on the device. Parakeet by default, or one of three Whisper sizes. Your personal dictionary is fed to the recognizer as a bias prompt before it decodes, not only applied afterwards. See Models and device fit.
• Clean up, deterministically. Seven ordered stages of plain code — no model — that remove hesitations, resolve spoken self-corrections, expand spoken punctuation, digitise numbers, build lists, and capitalise. See Cleanup and polish.
• Polish, optionally. A small on-device language model refines the cleaned text, at one of four strengths, with guards that throw its answer away and keep the deterministic text if it strays too far from what you said.
• Insert. The accessibility service splices the text in at your cursor. If it cannot reach the field, it falls back to the clipboard and tells you nothing was lost. See Text insertion and permissions.

Where to go next

Nine pages, each answering a different question. The sidebar carries the same list on every docs page.

How to read them, and how they can be wrong

Every number on these pages — a threshold in milliseconds, a model size, a default — was taken from the source rather than from memory or from the app's own marketing copy. That makes them accurate about the code in the repository at the time they were written, which is not the same as accurate about the build on your phone. If a figure here disagrees with what the app tells you, believe the app, and please open an issue so the page gets fixed.

There is no separate FAQ page. Questions live at the bottom of whichever page owns the subject, which is where you are already looking when you have them.

Every page here is also served as Markdown at the same address with a .md extension — this one is at /docs/index.md. Both are generated from the same source, so they cannot disagree with each other.