SayItFlow

100% on-device push-to-talk voice dictation for macOS.

Hold a key → speak → release.
Clean text appears instantly in any app.

Runs entirely on-device with Desert Ant Voz on the Apple Neural Engine + S1-mini.
Zero cloud. Near-zero idle footprint. Your audio never leaves your Mac.

---

Why most “local” dictation apps still feel heavy

Other on-device tools keep large models loaded 24/7 (often 1.2–2.5 GB of RAM).
They drain battery and keep your Mac warm even when you’re not speaking.

SayItFlow is designed to disappear when you’re not using it.

After 2 minutes of silence it cleanly unloads the models and drops to ~15–30 MB with 0% CPU/GPU.
When you press the hotkey again, the models stream back in while you speak — so by the time you release the key, everything is already ready.

---

What it does

• Hold Spacebar (or Fn / custom shortcut) anywhere
• Speak naturally
• Release → polished text is typed into the focused app (VS Code, Slack, Chrome, Terminal, Notes, etc.)

Key strengths

• Private — Audio never leaves your machine
• Fast — Voz on the Apple Neural Engine (up to 290× real-time)
• Light — Native Swift menu-bar app with intelligent model unloading
• Multilingual — 25 languages
• Clean English — S1-mini removes fillers, fixes punctuation & numbers

---

Requirements

• macOS 14.0+ (Sonoma or Sequoia)
• Apple Silicon (M1 / M2 / M3 / M4 or newer)

---

Install

Download

Grab the latest SayItFlow.dmg from the Releases page, drag it to /Applications, and launch.

On first run grant:

1. Microphone
2. Accessibility
3. Input Monitoring

Build from source

git clone git@github.com:innovatorved/sayitflow.git
cd sayitflow
make dmg

---

Usage

---

Architecture

Mic → Desert Ant Voz (Core ML / Apple Neural Engine) → S1-mini → Injected text

1. Speech recognition — Desert Ant Voz (467 MB, 25 languages, runs fully on the Neural Engine)
2. Text cleanup — S1-mini by Superwhisper polishes English (fillers, punctuation, numbers). Non-English is left untouched.
3. Hotkey engine — Low-level CGEventTap (cleanly distinguishes taps vs long holds)
4. Text insertion — Accessibility API + keystroke fallback
5. UI — Minimal monochromatic menu-bar design

Extra engineering for smoothness:

• Hardware-accelerated audio metering (Accelerate / vDSP)
• Pre-allocated 120-second ring buffer (zero mid-speech allocations)
• Cached regexes & formatters

---

License & Attribution

• SayItFlow - MIT © innovatorved
• Desert Ant Voz - Desert Ant Labs Source-Available License 1.0 (free under 100k MAUs)
• S1-mini - Apache 2.0 + naming clause © Superwhisper (must be credited as “S1-mini by Superwhisper”)

---

Built for people who want local voice input that doesn’t punish the rest of the machine.