Using On-Device Dictation

Why On-Device

The system keyboard's mic button — the one Apple shows next to the spacebar — uses standard iOS dictation, which has historically had server-side and learning components. For most people that's fine. For a security-focused vault, it isn't.

UltraLocked's mic button (in the editor's formatting toolbar) routes audio through Apple's on-device speech recognition. On iOS 26+ it uses SpeechAnalyzer with the long-form dictation preset; on earlier versions it uses SFSpeechRecognizer with requiresOnDeviceRecognition = true. Audio buffers and transcribed text never leave the device. We've verified this with airplane-mode testing and network proxy capture during active sessions.

Using Dictation

1. Open a note (tap any existing note, or create a new one via the Vault tab's floating +).
2. Place the cursor where you want text to land.
3. Tap the microphone icon on the right side of the formatting toolbar.
4. The first time you use it, iOS prompts for Microphone and Speech Recognition permissions. Both are required.
5. Start speaking. Words appear at the cursor as you talk. The mic icon turns red and pulses, and the bottom of the editor shows a red "Listening · tap mic to stop" indicator so you always know when the mic is hot.
6. Tap the microphone again to stop.

Text in flight (the part of a sentence the recognizer hasn't finalized yet) shows with a faint underline. Once it's finalized — usually within a moment of you finishing a phrase — the underline goes away and that text becomes a normal part of the document.

Auto-save: dictated text rides the same auto-save path as typed text, so it's encrypted and written to disk within a couple of seconds of being finalized. If the app is force-quit mid-dictation, the most recent finalized words are already saved.

Auto-stop: if there's no speech for 30 seconds, dictation stops on its own.

When Dictation Is Unavailable

Some locales don't support on-device recognition on every device. When that's the case, tapping the mic does nothing visible (the recognizer simply reports "unsupported"). It will not silently fall back to cloud-based transcription — that's a deliberate choice to preserve the on-device privacy guarantee.

If your locale isn't supported, you have a few options:

• Switch your iOS region or keyboard language temporarily (tap the globe icon on the keyboard) to one with on-device support.
• Type instead.
• Wait for Apple to expand on-device locale coverage; the architecture supports new locales as Apple ships them.

Pauses and Long Sessions

You can pause for several seconds mid-thought without losing your place — dictation continues seamlessly when you pick back up, and we handle Apple's per-task length limits internally so long writing sessions don't get clipped or reset.

What's Logged

Nothing about the audio or transcription is logged. We don't write transcribed text to Console.app, sysdiagnose, or any other system log. The state changes (started / stopped / errored) are logged at a coarse level for debugging, but the content is never recorded anywhere.

Audio Buffers

After each transcription ends, audio buffers are zeroed via the secure-memory utility before being released. The audio engine is fully torn down, the audio session is deactivated, and any cached recognizer state is dropped. There's no replay or audit trail.