In the open
What's shipped, and what's next
Nasar Flow is built in the open, so here's exactly where things stand. No fixed dates, just an honest status: this is a solo indie project, built without a Mac, shipped through GitHub Actions.
Shipped
Live in the beta todayRuns on whisper.cpp. Nothing is ever uploaded.
The default model, tuned on real Singlish speech.
No longer just shared Singlish vocabulary; its own model trained on real Malaysian speech, covering standard and everyday colloquial Malay.
Colloquial (Amiyah) Arabic, including common Islamic phrases.
Detects which language you're in and picks the right model.
Invite-only, automatically, straight from this site.
Turn Flow on once, then dictate straight into WhatsApp, Messages, Notes or Mail. Stop talking and it types itself; ABC keys for quick fixes; undo and put back.
Hold the mic: new line, delete that, full stop, capitalise that, "change X to Y", "spell A-L-Y". English, Malay and Arabic, never mistaken for dictation.
Raw / Light / Full dial, numbers and prices written the way you'd type them, and styles that read the field (Chat, Email, Notes, Exact). Never strips lah, leh, lor.
Written forms with many spoken forms, whole-word and case rules, CSV import and export, and spellings learned from the corrections you make by hand.
Recent audio kept on-device; failed dictations retried from History with any model. Clipboard fallback that expires.
A proper voice-activity model (Silero) finds your speech before transcription, so silence never becomes invented words. A hallucination filter catches the rest.
"Dictate with Nasar Flow" from a physical button, Back Tap or Siri. Three-screen first run.
In progress
Being actively worked onThe current model is workable but genuinely imperfect. Improving it is the top priority.
Today it's rule-based keyword matching. A trained classifier is next.
A Live Activity in the Dynamic Island and on the Lock Screen while the mic is open, with a one-tap Off. Plus a Lock Screen widget and Control Center buttons. Built and tested; ships in the next build.
The command table takes standard Malay and Arabic today. Everyday phrasings are being collected from native speakers rather than guessed.
Planned
On the list, not startedA small language model, entirely on the phone, for the "Full" cleanup notch and edits like "make this a list". Only if it stays under a second.
Measured error rates per dialect, including a code-switching test set, updated with every model change.
Model downloads should survive a force-quit mid-transfer, not just a backgrounded app.
Once the beta has proven itself out with real feedback.
After iPhone
Beyond iPhone: Mac, Windows, Android
The plan is to bring the same on-device engine to the other places people actually type: a Mac and Windows app, and Android. Real research into what each one takes is done - genuinely doable, each a real new build in its own right, not a weekend port. Android has an early working prototype; Mac and Windows are not started. No dates - same honesty as the rest of this page.
A menu-bar helper, not an extension - so unlike iPhone, no workaround needed to reach the microphone directly.
Paired with the Mac build from one shared codebase, following a real open-source app doing the same thing today.
Early prototype exists: it records, transcribes on-device and has a first keyboard. Android's keyboard framework doesn't block microphone access the way iPhone's does, so it may end up simpler than the iPhone app. Not in any store yet.
Have a request?
See something missing?
Open an issue on GitHub, it's the fastest way to reach me directly.
Open an issue on GitHub