Parakeet V3 vs Whisper: 7x Faster, Better Accuracy (Benchmark)

Why We Switched the Default

Whisper is great, but it was designed as a general-purpose model. It handles 100+ languages, translates, generates timestamps — a Swiss Army knife. The trade-off is speed. For English dictation, where you just want words on screen fast, it's overkill.

With Whisper behind Fn-key system-wide dictation, a one-minute utterance took 3–5 seconds to appear after you let go of the key.

With Parakeet the same utterance appears in about a second or less, which is short enough that you stop noticing the wait.

How Fast Is Parakeet V3?

Here are two of our own runs — both engines on the same audio, on the same Mac:

On a smaller machine the gap holds. On our M5 Air, Parakeet ran a 17.5-minute meeting in 16.7 seconds, roughly 60× realtime.

What Makes Parakeet v3 So Fast

Whisper's encoder reads a 30-second window of audio at once. Its decoder then writes the transcript one token at a time, and it always expects to write something, even over silence.

Parakeet is a transducer. It compresses the audio 8× first, then predicts each token together with how long it lasts, so it can skip ahead instead of stepping through every frame. Over silence it emits a "blank" and moves on.

That is where both the speed and the quieter pauses come from. On an M4 Pro it gets through an hour of audio in under a minute, with fewer parameters than Whisper Large V3 Turbo and a lower English error rate.

Benchmarks: Parakeet v3 vs Whisper

Lower is better: Parakeet V3, less than half the size of Whisper Large V3, stays within a point of it on all three benchmarks, and the larger Seamless M4T v2 has the lowest error rates.

On the Hugging Face Open ASR Leaderboard, Parakeet v3 averages 6.32% English WER with 600M parameters — less than half of Whisper Large V3's 1.55B:

Lower WER = fewer errors. Higher RTFx = faster. Parakeet is ahead on both, and at 600M parameters it is the smallest of the three, so it runs well on Apple Silicon without much memory.

Multilingual WER: All 25 Languages

The leaderboard above covers English only. Here's the full picture — how the three models in the table below compare across all 25 languages Parakeet supports, measured on the FLEURS benchmark. Lower WER = fewer transcription errors. Best value between Large V3 and Parakeet is highlighted per row:

WER (%) on FLEURS. Whisper Small data from Radford et al.; Large V3 and Parakeet V3 data from NVIDIA Canary-1B-v2 paper.

Whisper Large V3 edges ahead on most individual languages, but Parakeet V3 is close on average (12.0% vs 12.6%), leads on seven of them, most clearly Maltese and Greek, and cuts Whisper Small's average error rate by about 60%. The practical advantage is the combination: Large V3-level multilingual accuracy at roughly 60× realtime on Apple Silicon, a smaller runtime footprint, and much more reliable silence handling in our dictation tests.

Why It Hallucinates Less on Silence

If you've used Whisper for dictation, you've probably seen it hallucinate during silence — repeating phrases, inventing words, or spitting out "Subtitles by Amara.org" from nowhere. This happens because Whisper's autoregressive decoder always expects to produce text, even when there's nothing to transcribe.

Parakeet's transducer architecture can emit blank outputs instead of forcing a text token. In our system-wide dictation tests, that makes it far less likely than Whisper to fill a pause with repeated or unrelated text. It is still a speech model, so “less likely” is the accurate claim, not “never.”

Languages Parakeet Supports

Parakeet v3 supports 25 languages: 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.

That covers most of Europe, but not Chinese, Japanese, Korean, Arabic, or Hindi. For Chinese, Japanese, Korean, and Cantonese, choose SenseVoice; for Arabic, Hindi, and other languages outside Parakeet's set, use Whisper Large V3 Turbo.

Which Local Model Should You Use?

Whisper Notes ships four local speech-to-text models on Mac — Parakeet V3, SenseVoice Small, Whisper Small and Whisper Large V3 Turbo — and the Mac app from our website adds a fifth, Qwen3-ASR 1.7B (Beta). This section is about the four. Choosing between them is simpler than it looks: the language you are transcribing decides it almost every time, and your hardware decides the rest. Nothing about the choice affects privacy — all of them run entirely on your device, offline. What changes is speed, accuracy, and how much disk and memory the model wants.

Here is the whole decision in one table.

All of these run entirely on-device in Whisper Notes, and you pick the engine in Settings: the iPhone app ships Parakeet V3, Whisper Small and SenseVoice Small; on Mac you also get Whisper Large V3 Turbo, and the Mac app from our website adds Qwen3-ASR 1.7B (Beta).

And the same four models on the specs that decide the trade — Qwen3-ASR 1.7B (Beta), the fifth model in the Mac app from our website, is not in this comparison:

English WER from the Open ASR Leaderboard; 25-language averages from the FLEURS table above. Speed and download figures come from our own tests on an Apple M4 Pro running the Mac build — figures differ on iPhone.

Not sure whether your language is on the list at all? The per-engine table on our supported-languages page for Whisper, Parakeet V3 and SenseVoice is generated from the app's own source, engine by engine — it answers "is my language supported, and by which model" directly. And if the table above pointed you somewhere other than Parakeet, each specialist has its own benchmark write-up: SenseVoice for Chinese, Japanese, Korean and Cantonese, and Whisper Large V3 Turbo for the 101-language long tail.

Same Lineup, Right-Sized per Device: iPhone vs Mac

One difference is worth stating plainly, because it's a deliberate decision rather than a gap. On Mac, Whisper Notes has these four models: Parakeet V3 as the default, plus SenseVoice Small, Whisper Small, and Whisper Large V3 Turbo. The Mac app from our website also has a fifth, Qwen3-ASR 1.7B (Beta), which is not on iPhone.

On iPhone, Parakeet V3 ships built into the app, with Whisper Small and SenseVoice Small as optional downloads. Whisper Large V3 Turbo is intentionally not offered on iPhone: at roughly a 1.6 GB download that wants about as much again in memory while transcribing, it isn't a practical thing to ask most iPhones to carry. We'd rather right-size the lineup per device than ship a model that technically installs and then makes the phone miserable.

You don't lose language coverage on iPhone: Whisper Small exposes the same 101-language list as Turbo in a fraction of the footprint. What you trade is accuracy on the hardest languages — the correct trade to make on a phone. If you need Turbo-level multilingual accuracy on a long file, run that job on the Mac.

Parakeet V3 is built in and selected; Whisper Small and Whisper Large V3 Turbo are optional downloads that also run locally.

What Each Model Is Bad At

Every one of these engines is a trade, and a recommendation is only useful if it says what you give up. Here is the honest version.

Whisper (both sizes) hallucinates on silence. Long pauses, room tone, and dead air can produce repeated phrases, invented sentences, or the notorious "Subtitles by Amara.org" — the decoder is built to always emit text, so silence gets filled with something. Whisper Notes runs Pyannote voice activity detection before transcription to strip silence, which reduces it a lot, but no amount of pre-processing makes the model immune. Whisper Large V3 Turbo is also the heaviest option in the app: roughly a 1.6 GB download and about the same again in memory while transcribing. Whisper Small is the cheap way into 101 languages, and its multilingual accuracy shows the price — 29.8% average WER across the 25 languages in the table above, versus 12.6% for Large V3.

Parakeet V3 has a hard 25-language ceiling. All of them are European. There's no partial credit outside that list: feed it Japanese and you don't get worse Japanese, you get the wrong tool. No Chinese, Japanese, Korean, Cantonese, Arabic, Hindi, Turkish, or Vietnamese. It's also weaker than Whisper Large V3 on a number of individual languages — Danish, Finnish, Swedish, Slovenian and Latvian in our FLEURS table — so "Parakeet wins Europe" is true on average, not on every row. And while it hallucinates far less on silence, "far less" is the claim we'll defend, not "never".

SenseVoice Small is a specialist, not a generalist. Six choices — English, Simplified and Traditional Chinese, Cantonese, Japanese, Korean — and nothing else. It's superb inside that set and irrelevant outside it. It replaced an earlier Qwen3-ASR model in the Mac app, and it is on iPhone too.

And some limits are shared by all four. The Mac app requires Apple Silicon and macOS 14 or later — Intel Macs are not supported. Models are downloaded inside the app rather than bundled with it, so the first use of a new engine costs disk space and one download. Very long recordings need real memory and take proportionally longer on older hardware. And none of these models do live captioning: Whisper Notes transcribes after the recording stops, not while you speak. If you need real-time captions, this is the wrong app, whichever model you pick.

What About Parakeet V2?

If you used V2 before, you might wonder how it compares. V2 was an English-only model — and its English accuracy is actually slightly better than V3 (6.05% vs 6.32% WER). V3 trades that tiny margin for 25-language support. Both are significantly more accurate than Whisper.

In short: if you only need English, both V2 and V3 are great. V3 is the default in Whisper Notes because multilingual support matters to most users — and the English accuracy difference is negligible.

What About WhisperKit?

WhisperKit is Argmax's open-source Swift framework for running Whisper models on Apple devices. It's a developer toolkit, not an end-user app — and it runs Whisper, not Parakeet (which is NVIDIA's model family). If you want WhisperKit-style on-device transcription without writing Swift yourself, Whisper Notes packages the same idea as a ready-made app: Parakeet V3, Whisper (Large V3 Turbo on Mac, Small on iPhone), and SenseVoice, all running locally on Mac and iPhone.

Comparing all the local options? Every on-device speech-to-text model — Whisper variants, Parakeet V3, SenseVoice, and Voxtral — is compared side by side on our Whisper models comparison page. New to Whisper itself? Start with the Whisper Transcription Guide — what the model is, every way to run it, and what it costs.

Frequently Asked Questions

Try It

Parakeet V3 is the default engine in Whisper Notes on iPhone and Mac.

Questions or feedback? Email support@whispernotes.app.

Download for iOS Download for macOS