How on-device dictation actually works

What happens between the moment you speak and the moment text appears - no marketing, just the pipeline.

Written by the Bellow team

The whole path, end to end

When you hold the key and talk, this is the sequence:

1. Capture. The mic is opened and audio is buffered in memory - never written to disk.
2. Resample. Audio is converted to 16 kHz mono, the format speech models expect.
3. Transcribe. On release, the buffer goes to Whisper Large-v3 Turbo running on the Apple Neural Engine.
4. Clean. Non-speech markers the model emits are stripped.
5. Paste. The text is put on the clipboard and a synthetic ⌘V is posted to whatever app has focus.
6. Record. If history is on, the text is encrypted and stored locally. The audio is already gone.

No step involves a network request. There's no API key in the app, and nothing to configure.

What the Neural Engine is doing

Every Apple Silicon Mac has three compute units: the CPU, the GPU, and the Neural Engine - silicon dedicated to the matrix maths that machine-learning models are made of. It's fast at that specific job and uses far less power than doing it on the CPU.

Bellow uses WhisperKit, which compiles Whisper into Core ML so it runs on that hardware. That's why local transcription is quick enough to feel instant rather than feeling like a compromise.

Why pasting works in every app

Bellow doesn't integrate with individual apps. It writes to the system clipboard and synthesises a ⌘V keystroke into whatever currently has keyboard focus. Because that's the same mechanism you use yourself, it works in Slack, Mail, Notes, browsers, editors, terminals - anything that accepts a paste.

That's also why Bellow needs Accessibility permission: observing your trigger key globally and posting a keystroke both require it. It's the same capability launchers like Alfred and Raycast use.

Why it isn't on the Mac App Store

Watching for a hotkey across the whole system needs a CGEventTap, and App Store apps run in a sandbox that can't install one. Any dictation tool with a global push-to-talk key is distributed outside the App Store for this reason.

Bellow is signed with a Developer ID and notarized by Apple, which is the same trust check macOS applies to any direct-download app - Apple scans the build and issues a ticket that's stapled to the download.

Where your data lives

• Audio: memory only, discarded after transcription.
• The model: inside the app bundle. Read-only, never updated over the network.
• Transcripts: encrypted locally with AES-GCM; the key sits in your Keychain, device-only, never in iCloud.
• Settings: standard macOS preferences on your machine.

Nothing in that list has a server component, which is what lets the app run with no network permissions at all.