AI cleanup

Dictation already removes filler words on every recording. AI cleanup goes further: it follows your spoken edits and applies a writing style. Say “we should meet tomorrow at ten, actually no, eleven, and go over the budget” and get back “We should meet tomorrow at eleven and go over the budget.” The “actually no, ten” is gone because the model understood you were correcting yourself. Cleanup is off until you turn it on. Open the AI cleanup page from the sidebar and flip the switch at the top.

​

Shortcut

You choose how cleanup runs: Switch between the two on the AI cleanup page under Shortcut. When cleanup has its own shortcut, the keys are: Click the keys on the AI cleanup page to change them, the same way you change any shortcut.

​

SpeakoFlow Mini

SpeakoFlow Mini is a small model trained for one job: turning what you said into clean text. It removes filler words, fixes grammar and punctuation, and follows spoken edits, so “scratch that” or “actually, eleven” does what you meant instead of being typed out. It is a 795 MB download, runs on your computer, and handles English. Pick it on the AI cleanup page under Cleanup model, or from the Models page. When SpeakoFlow Mini is the cleanup model, the app shows a notice near the prompt editor: “This model is fine-tuned for English dictation cleanup, so it works best on the recommended prompt with no style.” The recommended prompt ships with the app, and you can restore it any time.

​

Other models

Any local or cloud model can do cleanup instead of SpeakoFlow Mini. Pick one under Cleanup model on the AI cleanup page. The same providers available for the assistant work here: your own downloaded model, Ollama, LM Studio, or a cloud provider with your API key. See Providers for the full list. On an Apple silicon Mac, Apple Intelligence also appears. It runs on-device with no key and no network, and needs macOS Tahoe 26.0 or later with Apple Intelligence enabled in System Settings. If no cleanup model is set up, cleanup borrows your assistant’s model automatically. Once the assistant works, cleanup works, with nothing extra to configure.

​

Cleanup instructions

Two prompts ship with the app: Open the prompt editor from the Cleanup instructions row on the AI cleanup page. You can edit either prompt, create your own, switch between them, and restore a shipped prompt to its original text with Restore recommended. Your transcribed text is sent to the model separately, so just write instructions.

​

Writing style

A writing style changes how the cleaned text reads. It is layered on top of the cleanup prompt, so corrections stay the same and only the register changes. See Writing styles.

​

Timeout and fallback

AI cleanup gets 10 seconds by default. If the model takes longer, your raw transcription is pasted instead. The timeout covers the whole round trip, including the first model load if the engine has been idle. On the built-in local engine, a cold load counts against it, so allow more time if your first cleanup of a session keeps timing out. Change the timeout on the Models page under your cleanup model’s settings. Options are 5, 10, 15, 20, 30, and 60 seconds.

​

When cleanup does not run

Cleanup fails soft every time. Your raw transcription is pasted and the dictation completes normally. That happens when:

• No model is configured, and neither is the assistant’s.
• A cloud provider is selected but has no API key.
• No cleanup prompt is selected, or the selected one is empty.
• The model returns an error, an unreadable response, or nothing at all.
• The model runs past the timeout.
• The output is implausibly long or looks like the model narrating its plan instead of returning cleaned text.

History records both versions of a cleaned dictation, so you can compare the original transcription with the final text.

​

Model unload

The cleanup engine runs on its own process, separate from the assistant. After 15 minutes of inactivity it unloads the model to free memory. The next cleanup reloads it. Change this on the Models page under your cleanup model’s settings if your machine is tight on RAM.