Until now, Sona waited. You held the key, spoke, released — and a beat later your cleaned-up text appeared. That beat is gone.
Streaming mode sends your audio to the transcription engine in small chunks while you talk. Partial text shows up in the HUD almost immediately and refines itself as more context arrives — punctuation, casing and corrections settle in real time.
Naive streaming gives you a jittery wall of half-words. We rebuilt the pipeline around partial results: every interim transcript is provisional, only finalized segments are committed, and the final pass still runs the same cleanup you already trust. The text you watch forming is a preview; the text we insert is the finished version.
On the desktop side that meant decoupling capture from transcription — audio now flows out frame by frame instead of as one buffer at the end — while keeping the classic batch path intact as a fallback. If the stream ever drops, you never notice: Sona just finishes the dictation the old way.
Streaming shines for longer thoughts — a paragraph of email, a spoken prompt, a whole chapter. For a three-word reply, batch is already instant. Toggle it in Settings; it stays off by default while we widen language coverage.
Hold the key, start talking, and watch the sentence assemble itself.