Whisper vs Parakeet

We ship both, so this is not a sales pitch

Most "X vs Y" articles are written by someone selling X. This one is different. We ship both engines in the same app. You pick the one you want from a dropdown, download it once, and switch any time.

That means I have no reason to talk you into the wrong one. If Parakeet is right for you, I want you using Parakeet. If you need Whisper, I want you on Whisper. The app does not pick for you and neither will I. I'll just lay out what each one is actually good at.

Both run fully offline. Both are free for every signed-in user. No card, no per-minute cloud bill, no paying extra for the second engine.

Parakeet is the fast one

Parakeet is NVIDIA's TDT model. That stands for Token-and-Duration Transducer, a FastConformer encoder bolted to a decoder that predicts both the words and how long each one lasts. You do not need to remember any of that. The part that matters is in the next sentence.

It is 5 to 10 times faster than Whisper on a plain CPU. No GPU required, no waiting. You release the hotkey and the text is basically already there.

The shipped model is Parakeet v3, about 600 MB on disk, and it covers English plus 24 other European languages, 25 in total. The quality is good. Not Whisper-large good, but more than enough for everyday dictation: emails, notes, messages, the stuff most of us type all day.

What Parakeet does not do: it will not translate your speech into English, and it does not take custom vocabulary or hotword hints. If you dictate in French, you get French text. Which is exactly what most French speakers want, so this is only a problem if translation is the whole reason you showed up.

Whisper is the thorough one

Whisper is OpenAI's open model, and inside our app it comes in eight flavors. Four English-only and four multilingual, from a roughly 140 MB Base model up to a roughly 3 GB Large v3. You pick the size that fits your machine and your patience.

The multilingual models cover 99 languages, not 25. That includes Chinese, Japanese, Korean, and Arabic, the ones Parakeet skips entirely. Whisper will also translate any of those into English in one step, which Parakeet cannot.

And Whisper gives you control Parakeet does not: custom vocabulary, beam-size settings, and hotword biasing so it stops mangling your colleague's name or your product's name. If you dictate technical terms all day, that control is the difference between clean text and a find-and-replace chore.

The cost is speed. On the same CPU, Whisper is slower, sometimes a lot slower on the bigger models. The boring truth is that most people picking Whisper are not picking accuracy over Parakeet so much as picking languages and control over speed.

The honest comparison, in one table

Here is the whole decision in seven rows. No dollar figures in it — both engines are free on the same tier.

Both are local, both are offline, both are free on the same tier. The table is not "good vs bad." It is "fast vs flexible."

Which one should you actually pick

This is the part most comparison articles dodge, so here it is straight.

For most English users, pick Parakeet. Speed wins daily. You dictate dozens of times a day, and the gap between "text appears instantly" and "text appears after a beat" is the gap between staying in your train of thought and losing it. Anything over about two seconds and your brain starts re-engaging with the app you were in, and you forget what you were saying. Parakeet being 5 to 10 times faster on a CPU is not a spec-sheet brag. It is the thing you feel a hundred times a day.

Pick Whisper when you need one of three things Parakeet can't give you: a language outside the 25 European ones (Chinese, Japanese, Korean, Arabic), translation into English, or custom-vocabulary and hotword control for names and jargon.

I am the kind of person who draws the architecture diagram before I install the runtime, so my instinct is always "use the more capable tool." That instinct is wrong here. The more capable tool is Whisper, but the right tool for a daily English dictation habit is usually the faster one. Capability you don't use is just download size.

How to switch between them

You do not have to commit. Open the model picker, choose Parakeet or any Whisper model, and it downloads once. Switch back whenever the work changes: Parakeet for the morning email batch, Whisper Large v3 the afternoon you're transcribing a Japanese interview.

Same hotkey, same overlay, same settings either way. On Windows the default is Ctrl+Space. On Mac it's a Command+Option push-to-talk chord, and it's remappable if it clashes with something. The engine is a dropdown, not a reinstall.

If you only remember one thing

Parakeet for speed and English. Whisper for 99 languages, translation, and control. You don't have to pick at install time, because both ship in the same app and switching is a dropdown, not a download decision you're stuck with.

If you want the longer picture on why running this on your own machine beats sending audio to a server, I wrote about that in offline speech-to-text. And if you're weighing us against a specific paid app, the superwhisper alternative piece covers that head-to-head.