# Local dictation and speech-to-text on Mac | Frog

Canonical page: https://frog.beffa.xyz/docs/dictation/
Updated: 2026-10-26

Open Dictation to turn microphone recordings into text through audio rules. Choose a speech model, a recording shortcut, and output behavior. Transcription can run locally on Apple silicon, with optional cleanup by a separate text model.

These guides describe Frog 1.0.42, the published download as of October 29, 2026. Download links lead to the latest published release.

## Set up an audio rule

The built-in Dictate shortcut is Option–Space, and shortcuts are editable. Record on the Dictation page starts a recording directly; the recording controls let you stop it. Escape cancels. Below the rules, Recording chooses the microphone, can mute Mac audio while recording, and takes an optional Cancel recording shortcut besides Escape.

Disabling Dictation in Features cancels active dictation and unregisters its rules’ shortcuts while preserving their settings.

1. Enable Dictation in Features if needed. Download a speech model in Models → Local models or add a compatible speech-to-text model in Models → Providers.
2. Open Dictation, add or edit an audio rule, and select its speech model. Set a transcription language where supported.
3. Choose hold-to-record or press-to-toggle recording and record your preferred shortcut.
4. Choose clipboard output or copy-and-paste output. Enable transcript cleanup only if you want a text model to rewrite the transcription.

## Local speech models and languages

Frog’s internal catalog includes Whisper variants using WhisperKit and Core ML, and Parakeet variants using FluidAudio and Core ML. Downloadable inference requires Apple silicon. Model weights are downloaded explicitly and are not bundled with the app.

Whisper multilingual models can detect language automatically or use a rule’s language hint. Parakeet v3 supports its own set of 25 European languages with automatic detection; Parakeet v2 is English-only. The selected model determines the language options and whether partial transcription is available.

## Keep your words or clean up the transcript

Without cleanup, the speech model’s transcript is the output. With cleanup, Frog applies the rule’s instructions using its selected text model. For example, a cleanup instruction can remove filler words and fix punctuation while preserving your meaning.

Cleanup adds a second processing step. A local speech model plus a local text model keeps both steps on your Mac. Cloud text cleanup sends the transcript and instructions to that provider. If cleanup fails, Frog preserves the original transcript on the clipboard.

Copy-and-paste output attempts insertion into the currently focused field after transcription, so you can change input fields while recording. Password fields are excluded when identified by Accessibility. The clipboard remains useful if the destination app cannot accept automatic insertion.

## When does Frog record?

Recording starts when you invoke an audio rule or Record and stops automatically after ten minutes. Local transcription keeps microphone audio in process memory and discards it after completion or cancellation. If transcription fails in a way a second attempt can fix, the recording stays in memory until you retry or discard it in the popup. Choosing an external speech model sends the recording to that provider. Audio files are not added to Frog’s history.

Saved result history is a separate opt-in setting in Settings → History. Clipboard collection is enabled independently in Features, and copied text appears on the Clipboard page.

## Related guides

- [Local models](https://frog.beffa.xyz/docs/local-models/)
- [Model providers](https://frog.beffa.xyz/docs/providers/)
- [Privacy and offline use](https://frog.beffa.xyz/docs/privacy/)

## Source documentation

- [Audio rule settings](https://github.com/t1llo/frog/blob/main/docs/providers.md)
- [Speech model capabilities](https://github.com/t1llo/frog/blob/main/docs/local-model-sources.md)