Voxpen Now Runs NVIDIA Parakeet v3 on the Neural Engine

Since day one, Voxpen has run OpenAI's Whisper models locally on the Apple Neural Engine. That was the whole pitch: dictation that never leaves your Mac. Whisper is still excellent — but the on-device speech world moved fast this year, and one model has clearly pulled ahead for real-time dictation: NVIDIA's Parakeet.

Starting with v1.4.0, Voxpen ships Parakeet TDT 0.6B v3 as a selectable engine, alongside the full Whisper lineup. Here's what that means and why we did it.

What is Parakeet?

Parakeet is a family of open speech-recognition models from NVIDIA. The v3 release is a 600-million-parameter model that covers English plus 24 other European languages. On accuracy leaderboards it matches or beats Whisper's large models on English — while being roughly 10x faster.

The "TDT" part (Token-and-Duration Transducer) is the interesting bit for dictation. Unlike Whisper, which decodes audio in 30-second windows and can hallucinate text during silence, Parakeet's architecture predicts tokens and their durations directly. In practice: no more phantom "thank you for watching" when you pause mid-thought.

The numbers

On Apple Silicon, running on the Neural Engine via CoreML, Parakeet v3 transcribes at up to 110x realtime — a full minute of speech in about half a second. In our own testing, a typical short dictation transcribes in roughly a tenth of a second. Press the hotkey, speak, release — the text is pasted before your finger leaves the key.

And because it runs on the Neural Engine rather than the CPU or GPU, power draw stays around 0.3W. No fan spin, no battery drain, no slowdown in whatever app you're working in.

How we integrated it

Parakeet runs in Voxpen through FluidAudio, an open-source Swift SDK that compiles the model to CoreML and schedules it on the Neural Engine. It sits behind the same engine interface as our WhisperKit integration, so everything you already rely on — voice activity detection, Polish Mode, auto-paste — works identically regardless of which engine you pick.

Nothing about our privacy model changes. The model downloads once (~500 MB), and from then on every byte of audio is processed on your device. Offline, as always.

Whisper isn't going anywhere

Parakeet covers 25 European languages. Whisper covers 99. If you dictate in Japanese, Hindi, or Arabic, Whisper remains your engine — and we've upgraded that side too. The new large-v3-turbo model joins the lineup: Whisper's multilingual flagship with a pruned decoder that runs about 8x faster than large-v3 at near-identical accuracy.

There's also a new Models tab in Settings where you can download, switch, and delete models freely. Switching engines now takes effect instantly — no restart required.

Which model should you pick?

Our honest guidance:

• Parakeet v3 — if you dictate in English or a European language. It's the fastest and doesn't hallucinate on silence. This will become our default for English users.
• distil-large-v3 — the balanced Whisper default if you want maximum English accuracy from the Whisper family.
• large-v3-turbo — if you regularly dictate in languages outside Parakeet's 25.
• large-v3 — heavy accents, difficult audio, rare languages. Slowest, most thorough.

Different tools for different jobs — and unlike every other dictation app we know of, you get to choose.