Clean up dictation. Keep your voice.

Recognition hears words. Cleanup tidies them.

Speech recognition turns sound into a transcript. Cleanup takes that transcript and edits it. Keeping those stages separate lets you choose whether your words need only a small rule-based adjustment or a model that can interpret a false start.

Echo transcribes speech on your Mac. After that, the mode in Settings → Polish controls the cleanup. You can start with Off and add a model later; an AI subscription is not required to use the app.

Choose how much help you want

Off: a small, predictable pass

Off removes selected standalone fillers, including “um,” “uh,” “erm,” and “euh,” and applies your dictionary. It does not use a language model. Choose it when you want to inspect the transcript closely or keep the fewest moving parts in an offline setup.

Apple: cleanup on your Mac

Apple mode uses Apple’s on-device model to tidy punctuation and capitalisation. It depends on Apple Intelligence being available and prepared on your Mac. Echo checks the result and can keep your original, rule-cleaned text when the model’s response looks wrong.

AI: the provider you choose

AI mode asks your selected tool to handle repetitions, self-corrections, and misheard words as well as punctuation. Echo supports installed Claude and Codex tools, a local model through Ollama or LM Studio, and a custom command. Cloud tools use your own account and its usage terms; a model running on your Mac can keep the cleanup local.

Open Settings → Polish, choose the provider, and use the available status and Test controls to confirm it is ready. The offline guide covers which models to prepare before disconnecting.

Give cleanup a clear job

Echo tells the model to keep your wording, tone, and language. It should not answer a dictated question, translate the text, add an introduction, or turn a short thought into a summary. If you say “Can you send me the report?”, you want that sentence inserted—not an answer about whether a report can be sent.

The result checks look for problems such as an added preamble, large changes, missing phrases, or an invented day. When a result fails those checks, Echo uses your rule-cleaned words instead. If a provider is slow or unavailable, Echo can fall back to Apple-cleaned or raw text.

These checks reduce unwanted rewriting; they do not establish that every name, number, or fact is correct. Keep the final review with the person writing the message.

Test a mode with your own sentences

Use three examples: a clean sentence that should barely change, a sentence with a correction, and a sentence with a name from your work. Try a second language too if you use one. Read the output for meaning before judging how polished it sounds.

For recurring names, use Settings → Dictionary rather than expecting a model to guess. If cleanup keeps changing a phrase you want to preserve, compare the result with Off. That separates a recognition problem from an editing problem.

In a daily workflow, start with a short email paragraph. Pick the mode that produces the least editing for your own writing, and remember that selecting a cloud provider sends the transcript to that provider. Speech audio remains on your Mac; see the privacy details for the full distinction.