Parakeet: real-time streaming dictation on a Mac (words ~0.3s behind your voice)

Most on-device dictation works in one shot: you hold a key, talk, release, and the app decodes the whole clip at once. Parakeet is different by design. It’s a streaming engine, so text appears while you’re still speaking — partials trail your voice by about a third of a second, already carrying punctuation and capitalization — and the final result lands roughly 50 milliseconds after you let go of the hotkey. It’s a new engine you can pick in OpenWhisp’s settings, and it runs entirely on your Mac.

Streaming vs. batch: why the latency feels different

A batch engine (like WhisperKit, the engine behind offline dictation on a Mac) buffers your whole utterance and transcribes it after you stop. That’s accurate and fully local, but there’s a decode pause on release — the app has to think about the clip before the words appear.

A streaming engine is architecturally continuous: it’s transcribing as the audio arrives, so you watch the sentence build in real time. With Parakeet the partials trail your voice by about 0.3 seconds, and because it has already been keeping up, pressing stop doesn’t trigger a fresh decode of the whole clip — the final text settles in about 50 milliseconds. There’s no end-of-dictation wait, which is the part that makes it feel immediate.

Four streaming model variants

Parakeet isn’t one model but a small family, so you can trade speed for accuracy or coverage for size:

• Realtime English — the unified low-latency model tuned for live dictation.
• Best-accuracy English — when you want the cleanest English transcript.
• Ultra-light 120M — the smallest, lightest variant.
• Multilingual — roughly 40 languages, auto-detected (you can also give it a language hint).

It also powers files, meetings, and history

Parakeet isn’t only for live dictation. A separate 25-language batch model backs the file-transcription side too: the batch file queue, watch folders, meeting capture, and re-transcribing anything already in your history all run on Parakeet when you choose it. So the engine you dictate with is the same one that turns your recordings and meetings into text.

On-device and private, like the rest of OpenWhisp

Parakeet is built on the Apache-2.0 FluidAudio SDK and runs NVIDIA Parakeet models through CoreML — entirely on your Mac. Nothing is uploaded, which keeps it in line with OpenWhisp’s on-device, no-cloud approach. It’s transcription only: there’s no translation step, so your words come back in the language you spoke them.

The default engine — but not the only one

Parakeet tested well enough that it’s now the default, recommended engine: a fresh install downloads and uses it from the very first launch. If you were already using OpenWhisp, an update leaves you on whatever engine you’d chosen — it never swaps a working engine out from under you. Either way it’s not a lock-in: WhisperKit is still there in Settings › Models, as are whisper.cpp and Apple Speech, and the FluidAudio model repositories show up in Settings › Storage with their sizes and a one-click delete, so you only keep the variants you actually use. Weighing the options? The Mac dictation apps guide and the Wispr Flow alternative writeup put the trade-offs in context.

Bonus: agents can stop the moment you finish speaking

Parakeet also ships an end-of-utterance model that emits a genuine “the speaker finished” signal. When a coding agent hands you the mic, OpenWhisp can use that signal to end the session right after your utterance settles, instead of waiting out a silence timer. This is experimental and off by default, it applies only to agent-requested dictations, and it’s inert on every other engine — your own hotkey dictations are never auto-stopped. It pairs with talking to coding agents by voice.

Try it

Parakeet ships in OpenWhisp, a free, open-source dictation app for Apple Silicon Macs. Get it from the OpenWhisp site or GitHub, then pick Parakeet in Settings › Models.