Voice typing for Mac.
Keep your flow.

A shortcut to your first draft.

ChatterKey lives in the Mac menu bar. You can dictate into supported editors without starting a separate document, or select a passage and speak an editing instruction.

• Flexible recording: Fn hold-to-talk, Fn double-tap hands-free, manual Stop and Escape to cancel. Other configurable shortcuts remain hold-to-talk.
• Writing modes: professional, concise, technical, bullets, translation, verbatim and other selectable modes. Add vocabulary, snippets and custom instructions.
• Voice-edit review: compare the original and proposal, acknowledge protected-value changes, then Apply, Copy or Discard.
• Local tools: optional rough live transcript, optional history, usage dashboard and editable provider cost estimates.

Each live dictation or voice-edit attempt uses one request to the selected audio-capable model. No hidden transcription/polishing pair or automatic retry. Choosing Retry explicitly makes another request and may incur another charge.

New in Mac 6.0.0: bring your recordings.

Turn a recorded meeting, lecture or voice memo into a Raw transcript, detailed Notes or a short Summary. The separate import window keeps the transcript available alongside notes and summaries for review, Copy and Save.

1. Open Import Audio from the Mac menu and choose an MP3, M4A, WAV, AIFF or CAF file supported by macOS. File selection only prepares audio locally.
2. Choose an output format and check the request count. Generate sends prepared audio to your selected provider; provider fees apply.
3. Review the result and source transcript, then Copy or Save. Imported results are never automatically pasted or submitted, and stay outside dictation history and usage counters.

Long recordings, one selected model.

• Up to and including 30 minutes: one model request for any output format.
• Over 30 minutes: local 15-minute chunks, transcribed sequentially by the same selected model/provider, then joined in chronological order without summarising.
• Raw transcript: the merged transcript, with no extra request. Notes or Summary: one additional request using the complete transcript, only after every chunk succeeds—not summaries of individual chunks.

A 50-minute-48-second recording produces four chunks: 15 + 15 + 15 + 5:48 minutes. That means four requests for Raw transcript, or five for Notes/Summary. The app shows this count before Generate.

Clear progress and explicit recovery.

Follow chunk progress, cancel, or explicitly retry failed work. Completed chunks remain in memory and are not sent again on resume; a failed final notes/summary step retries only that step. Keep the same connection and model to resume. No automatic retry, repair call, fallback or provider/model switch is made.

Copy or Save before closing. Normal close/quit asks before discarding active work, completed chunks or results, but checkpoints do not survive a restart. Clear or a new selection also discards them. Cancellation cannot recall uploaded data or reverse provider charges.

Limits and privacy.

Local limits: 256 MiB source file, 4 hours and 16 chunks; at most 440 MiB temporary WAVs, 4 MiB merged transcript and 8 MiB per response. Requests run sequentially with 300-second timeouts. Provider audio, context or output limits may be lower; final Notes/Summary requires structured-output support.

Notes instructions ask to preserve substantive details, names, numbers, examples, explanations and qualifications without adding unsupported facts. AI can still mishear or omit content—there is no zero-information-loss guarantee. Review the full transcript against the original recording when accuracy matters.

Choosing a file does not upload it. Generate sends audio to your selected provider; final long-recording Notes/Summary also sends the complete merged transcript. Temporary files are cleaned up after active workers release them on Clear, new selection, close or exit; the original recording is never modified. Smaller MP3 uploads do not bypass model limits or promise lower audio-token usage.

Also in version 6: a compact 320-point menu-bar popup, native clear glass on macOS 26+, and older-macOS/Reduce Transparency fallbacks. Import-in-progress status prevents starting dictation while an import is busy.

Validation: local audio preparation and mocked-provider regressions cover chunking, ordering, cancellation and explicit recovery. Live long-recording provider quality and acceptance remain unverified; an earlier repeated chunk failure has not been confirmed resolved in a live run.

How to set up voice typing on Mac.

1. Check compatibility and the package. Use an Apple Silicon Mac on macOS 14 or later. Get the ZIP and checksum from the Mac download section. This release is not an Intel/universal build.
2. Install the app. Quit an older ChatterKey copy, extract the ZIP, move ChatterKey.app to Applications and open it. Existing settings and Keychain accounts are retained when updating normally.
3. Connect your provider. Open Settings → AI Provider. Choose Google Direct for your Gemini API key, or OpenRouter for its own key. Save an audio-capable model available to that account. Voice processing requires internet and provider access.
4. Allow the permissions you need. Microphone captures speech; Accessibility supports the shortcut, selection and paste. Speech Recognition is optional for the on-device rough preview.
5. Try a disposable draft. Focus an editable note, hold Fn, say a short sentence, then release. For hands-free, double-tap Fn and press it again when finished. Review the inserted text before sending or saving.

Rewrite selected text with your voice.

Select the exact editable passage in the destination app—not text copied earlier—and keep the selection in place. Use the recording shortcut and say, for example, “Make this more concise without changing the dates.”

ChatterKey sends the selected text and instruction audio in one request, then shows a review. It highlights literal edits and warns when detected numbers, emails or web URLs change. Those warnings do not verify meaning or factual accuracy.

Why might Apply be unavailable?

The original app, field or selection may have changed, or the editor may not expose enough Accessibility information. ChatterKey won’t guess a replacement target. Choose Copy & Close and paste the reviewed text yourself.

A terminal emulator’s screen selection is not necessarily the selection inside its editor buffer. Read-only pages also can’t be replaced in place. Even when a paste event is dispatched, that is not confirmation that the destination accepted it. No voice action submits your message automatically.

Your provider. Local controls.

Google and OpenRouter keys stay in separate macOS Keychain accounts. Audio and instructions go directly to the selected provider; an explicit voice edit includes the selection. There is no ChatterKey-operated transcription server, but the provider’s own data policies apply.

Optional local transcript history is disabled by default. The dashboard stores local usage metadata and estimates, not exact provider billing. Temporary audio can be retained after a failure for explicit Retry and is cleared on the documented recovery/cleanup paths. See the full privacy policy for storage, clipboard and cleanup details.

The native Android voice keyboard shares processing rules, not Mac permissions, settings or API keys. It does not include the Mac history or cost dashboard.

If something doesn’t work.

The shortcut isn’t responding.

Check the configured shortcut and Microphone/Accessibility access in macOS System Settings. Recheck permissions after replacing the app. Fn supports both hold and double-tap; other shortcuts remain hold-to-talk.

The provider rejects the attempt.

Check that the key belongs to the selected connection, that the chosen model accepts audio, and that your account has access and billing/quota available. ChatterKey shows the error instead of silently changing providers. Retry is your choice.

An imported chunk keeps failing.

Keep the import window open to retain completed chunks. Check the displayed error before another billable attempt; don’t repeatedly retry a deterministic failure. The same model may have audio or output limits below the app’s local limits. No final notes are produced from an incomplete chunk set.

The text didn’t arrive in the editor.

Check that the original destination is still editable and focused. Use the available Copy recovery and paste manually. Test with non-sensitive sample text; don’t publish real drafts, API keys or device logs in issues.

See the macOS 5.0.0 release notes for tested scope. No clean-account, exhaustive cross-editor or real-provider accuracy/latency guarantee is made.