Whisper vs Parakeet on a Mac: which model to use

Use Parakeet V3 if you speak one of its 25 European languages: it is a 494 MB download and runs offline on your Mac. Use Whisper large-v3 turbo (1.5 GB, or 547 MB quantized) for any other language, since Whisper covers about 99. Sizes are from Nota’s model list as of 2026-09-27. On public benchmarks the two are close in accuracy, so language decides first.

What Parakeet and Whisper are

Whisper is OpenAI’s open speech model family, released in 2022 and updated since. It was trained on a very large multilingual set, which is why it covers so many languages. Sizes run from tiny (75 MB) to large-v3 (2.9 GB). Large-v3 turbo is a trimmed large model that keeps most of the accuracy at about half the size.

Parakeet TDT 0.6B V3 is NVIDIA’s open speech model, 0.6 billion parameters. V2 was English only. V3 added 24 more European languages. It is built for speed: it predicts words and how long they last in one pass, so it skips work that Whisper-style decoders do word by word.

On a Mac, both run through local runtimes (whisper.cpp for Whisper, Core ML for Parakeet). Neither needs an account or a key.

Languages: the deciding factor

Parakeet V3’s 25 languages, from the model card: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian.

If your language is on that list, Parakeet V3 is a good default. If it is not (Japanese, Chinese, Korean, Arabic, Hindi, Turkish, Vietnamese and most others), use Whisper. There is no setting that makes Parakeet understand a language it was not trained on.

Mixed languages in one sentence work best on a model that covers both. Parakeet handles English plus Russian or Ukrainian; for English plus Japanese, it has to be Whisper or a cloud model. More in dictating in two languages.

Size and memory on a Mac

Download sizes and languages as listed in Nota 2.14 on 2026-09-27. Memory is Nota’s own estimate of what the model needs while it runs.

Model	Download	Memory estimate (GB)	Languages	Good for
Parakeet V3	494 MB	0.8	25 European	Default for most European-language users
Parakeet V2	474 MB	0.8	English	English only
Whisper base	142 MB	0.5	About 99	Older or low-memory Macs, quick tests
Whisper large-v3 turbo q5	547 MB	1.0	About 99	Non-European languages on a small footprint
Whisper large-v3 turbo	1.5 GB	1.8	About 99	Non-European languages, best balance
Whisper large-v3	2.9 GB	3.9	About 99	Maximum Whisper accuracy, slowest

Quantized means the model's numbers are stored with fewer bits, so the file is smaller and loads faster, at a small cost in accuracy.

Large-v3 needs about 4 GB of free memory while it runs, by Nota’s estimate, so on an 8 GB Mac with a browser open, turbo q5 is the safer Whisper.

Accuracy: what the public benchmarks say, dated

Two primary sources, both checked 2026-09-27:

• English. On the Hugging Face Open ASR Leaderboard, Parakeet TDT 0.6B V3 averages 6.34% WER, as quoted on its model card. NVIDIA’s paper on the model (September 2025) lists Whisper large-v3 at 7.44% on the same benchmark set.
• Multilingual. The same paper reports Parakeet V3 at an average 9.7% WER across its European languages (FLEURS, CoVoST and MLS test sets), against 9.9% for Whisper large-v3.

So on NVIDIA’s own numbers, Parakeet V3 is slightly ahead in English and about even in its other languages, with well under half the parameters (0.6 billion against 1.55 billion). Whisper’s advantage is not accuracy, it is the other 70-odd languages.

Two cautions. Benchmarks are read speech and clean recordings, not you talking into a laptop mic in a café. And a gap of half a point of WER is smaller than the difference a good microphone makes.

Speed is the clearer difference. Parakeet is a smaller model with a decoder designed for speed, while Whisper large decodes word by word. We don’t publish our own timing numbers; try both on your Mac with the same sentence.

Which one Nota starts you on, and how to switch

Nota’s setup asks which languages you dictate in, then suggests a local model for them and downloads it when you choose Local. That is the whole setup: no key, no account, and your audio stays on your Mac. Parakeet V3 is one of the local models you can pick on the Models page.

To switch:

1. Open Models.
2. In the guide strip, set Languages and Where: This Mac. The recommended model moves to the top with a one-line reason.
3. Click Download on the model you want. Whisper models show “Optimizing for this Mac” for a moment after the download.
4. Pick it in your mode. Each mode can use a different model, so an English mode can stay on Parakeet while a Japanese mode uses Whisper.

Cloud models are there too, on your own key, if a local one is not good enough for your language. See local models and cloud models.