Speak, and it's written. Speak again, and it's rewritten.
Native iOS dictation. Tap the mic, speak, and the transcript lands on your clipboard — ready to paste anywhere. Then speak again within 30 seconds to rewrite it with your voice. Entirely on-device. No cloud. No accounts. No telemetry.
1. Tap the mic. From inside the app, the Dynamic Island, or your Jot keyboard.
2. Speak naturally. The pill shows amplitude in real time. No length limits. Cancel any time.
3. Transcript lands on your clipboard. Paste it into Messages, Notes, Slack — any app.
4. (Optional) Speak a follow-up. Within 30 seconds, say "make it friendlier" or "make it shorter" and Jot rewrites the previous transcript in place. No typing, no re-speaking from scratch.
On-device privacy. Audio is transcribed locally on the Apple Neural Engine using Parakeet TDT 0.6B v3 via FluidAudio. Nothing crosses the network. Run Little Snitch — you'll see silence during transcription.
Voice-driven follow-up. A 30-second window after every dictation turns your next utterance into a potential rewrite command. An on-device LLM (Apple Foundation Models) classifies command vs new dictation. A deterministic pre-classifier catches obvious command shapes ("change that to...", "make it...") without an LLM round-trip. No network, no leak.
Full-access keyboard. A custom keyboard extension with one-tap paste from your transcript history. Switch to the Jot keyboard from any app, insert a previous transcript without leaving.
Live Activity + Dynamic Island. Recording status, transcribing progress, and the 30-second follow-up countdown are visible from any screen on the system. Tap to dismiss the follow-up window without opening the app.
Transcript history. Every dictation is saved on-device (SwiftData). Your history stays with you; nothing syncs anywhere.
Honest UX. Chrome is monospaced, tracked caps, em-dash wraps. Status is a ghost ledger note, not a toast. The UI reads like a terminal log, not a setup wizard — because one more permission prompt is not a feature.
Chained follow-up as voice-driven editing. Most dictation apps hand you raw text. Jot lets you iterate on it by speaking — the 30-second window is an editing mode in disguise. "A second thought" becomes a clean rewrite, classified locally, applied to the previous transcript in place.
Ledger aesthetic. Inspired by terminal logs: monospaced fonts, tracked caps for system labels (READY, PROC, CLEAN, FOLLOW), em-dash wraps for placeholder voice (— no entries —), muted gray hierarchy, amber only for the recording dot.
Cold-load honesty. The first transcription after launching Jot takes ~19 seconds to load Parakeet into the Neural Engine. Instead of lying about it, Jot shows a ghost ledger note:
WARMING
— first transcription may take ~20s —
Reactive to actual model state. Vanishes instantly when the load completes. Subsequent recordings are ~100ms.
Core ML's ANE-compiled cache persists across app launches and device reboots (per WWDC23 session 10049). We're building a BGProcessingTask that wakes Jot periodically to keep that cache primed — so after the very first use, every cold launch becomes a warm load.
The dream: press the Action Button → mic starts instantly → no app visible → paste the transcript when you're done. Zero bounce.
What works today:
- iOS 18+
AudioRecordingIntentpromotes the intent'sperform()into the main app process NSSupportsLiveActivitiesdeclared, Live Activity starts cleanly- iOS recognizes Jot as actively recording; the audio session activates without foregrounding
What doesn't (yet): AVAudioEngine.start() fails with OSStatus 'what' (2003329396 — "invalid state") when the app is running-active-NotVisible. The same session configuration succeeds while visible. Same process. Same session ID. The only delta is process visibility.
Apple DTS thread 65604 (Quinn "The Eskimo!") confirms the rule:
"The audio background support only allows you to continue using an audio session that you created in the foreground; it does not allow you to start a session from the background."
The limit is at the AVAudioSession layer. AVAudioRecorder hits the same rule — we researched it. The Action Button path promotes your process and registers you as recording, but Apple doesn't grant a cold-start exception.
Current fallback: Action Button bound to DictateIntent (openAppWhenRun = true) briefly bounces the app, records, returns. It works; it's just not zero-bounce.
Research lanes we're watching:
BGContinuedProcessingTask(iOS 26+) — foreground-start, background-continue. Might allow a stealth foreground flicker that reads as zero-bounce.- Pre-established engine — keep a recording-ready session alive across app launches via the audio background mode.
- iOS releases — Apple may ease the restriction in a future version.
We're not shipping a zero-bounce Action Button until it actually works on the current platform. No vapor.
Swift 6 · SwiftUI + UIKit interop · SwiftData · FluidAudio (Parakeet TDT 0.6B v3) · Apple Foundation Models · ActivityKit · AVAudioEngine + AVAudioConverter (16 kHz mono Float32) · Custom keyboard extension with Full Access
- Apple Silicon iPhone (iPhone 15 Pro or later — Apple Intelligence + ANE required)
- iOS 26.0+
Jot is reproducible from XcodeGen specs so the Xcode project stays out of git.
brew install xcodegen
cd Jot/
xcodegen # regenerate the Xcode project from project.yml
open Jot.xcodeprojThen pick a real device (the Simulator doesn't have the Neural Engine, so Parakeet won't run), sign with your Apple ID team, and build.
- Jot for macOS — the desktop sibling: press a hotkey, speak, text appears at your cursor. Same transcription engine, different surface.
TBD