Parakeet vs Whisper: which speech model for Mac dictation?

Whisper, from OpenAI, is the multilingual choice: one family of models in several sizes that covers many languages. Parakeet, from NVIDIA, is a smaller family: version 2 is English only and version 3 covers 25 European languages. NVIDIA's model cards report low word error rates for both on public benchmarks. Neither is the best voice to text model for everyone. Pick by the language you speak, then test both on your own voice, because benchmark audio is not your microphone, accent or vocabulary.

Publisher note: SuprFlow publishes this guide and is one of the products discussed. Competitor details come from each company's own pages, checked on 1 October 2026. Prices and features change; confirm before you buy.

What are Whisper, Parakeet and WhisperKit?

Whisper is described in OpenAI's repository as "a general-purpose speech recognition model" that can perform "multilingual speech recognition, speech translation, and language identification." The accompanying paper by Radford and colleagues says it was trained on 680,000 hours of multilingual and multitask supervision.

Parakeet is a family of speech recognition models from NVIDIA. The parakeet-tdt-0.6b-v2 model card describes "a 600-million-parameter automatic speech recognition (ASR) model designed for high-quality English transcription." The v3 card says it extends v2 "by expanding language support from English to 25 European languages."

WhisperKit is not a third model. People search for Parakeet vs WhisperKit, but WhisperKit is software from Argmax for running Whisper models on Apple devices. Its repository, now named argmax-oss-swift, describes a collection of on-device inference frameworks. WhisperKit requires macOS 14 or later and Apple silicon, and downloads Whisper models converted for Apple's Core ML. Comparing WhisperKit with Parakeet is really comparing Whisper, run a particular way, with Parakeet.

How do the models compare on paper?

Sources: the Whisper README's model table, the large-v3 model card and the two NVIDIA model cards. Both Parakeet cards say the model outputs punctuation and capitalization. The v3 card lists its languages as Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian.

Two points follow. If you dictate in Hindi, Japanese, Arabic, Mandarin or any other language outside that list, Parakeet v3 does not list it, Whisper's large-v3 card does, and the comparison is over. And parameter counts are not download sizes: apps ship these models in converted and compressed forms, so check the size each app shows before you download.

What do the published accuracy numbers say?

Accuracy for speech models is reported as word error rate (WER): the share of words the model gets wrong against a reference transcript. Lower is better.

NVIDIA's model cards report these results on the eight English datasets used by the Hugging Face Open ASR Leaderboard:

For Whisper, OpenAI's README does not give a single average. It says "Whisper's performance varies widely depending on the language" and shows error rates by language on the Common Voice 15 and Fleurs datasets. The large-v3 model card adds that the models "exhibit disparate performance on different accents and dialects." To compare Whisper and Parakeet figures like for like, open the leaderboard and read the current rows yourself. We give no Whisper average here, because OpenAI's own pages do not state one.

Read all such numbers with three cautions:

1. They are benchmark datasets, not your voice. Look at the spread within one model: under 2% on clean audiobook recordings and over 11% on meeting recordings. Your accent, microphone, room and vocabulary will move the number more than the difference between two good models.
2. The Parakeet figures come from NVIDIA's own cards. The v3 card also notes that its multilingual error rates are calculated after removing punctuation and capitalization, so they say nothing about how well commas and capitals come out.
3. The model is not the app. A dictation app adds its own audio capture, silence handling, dictionary and formatting. Two apps using the same model can give you different text.

Whisper has a known weakness worth naming: the large-v3 card says its architecture makes it "prone to generating repetitive texts" and mentions hallucinations. Why words appear after silence covers what that looks like in practice.

Which Mac apps offer which model?

From each app's own pages:

If your goal is to try Parakeet and Whisper side by side on a Mac, an app that offers both is the practical route, and Handy and Spokenly let you do it without paying. SuprFlow 1.0.0 does not offer Parakeet, so it is the wrong tool for that comparison.

Which should you choose?

• You dictate in a language outside Parakeet v3's 25: Whisper.
• You switch between many languages, or want speech translated to English: Whisper, in a multilingual size other than turbo for translation.
• You dictate in English or a listed European language: try both. NVIDIA's reported error rates are low. Decide on your own test, not on the table above.
• You need a local speech to text model on a Mac with little memory: Whisper comes in smaller sizes (tiny, base, small), and the README's table shows them needing less memory and running faster than the large model. Check each app's stated requirements.
• You care most about names and jargon: the model matters less than whether the app has a dictionary. See Teach SuprFlow names and acronyms for how one app handles it.

A 15-minute test on your own voice

1. Install one app that offers both models, and download one of each.
2. Read the same fictional passage with each: "Dr Imani Okafor moved the 14 March review to 3:45 pm. The Q3 variance was 2.4 per cent, not 24. Ask Zbigniew to send the Kestrel-9 report."
3. Repeat in the room and with the microphone you normally use, then once with background noise.
4. Count wrong words, wrong numbers and wrong names for each model. Note how long each took to return text.
5. If you use a second language, repeat in that language.

Whichever model needs fewer fixes on your passage is the better model for you.

Questions people ask

Is Parakeet more accurate than Whisper?

On the English benchmark datasets in NVIDIA's model cards, Parakeet's reported error rates are low. That does not tell you which is more accurate on your voice. We have not run a controlled benchmark, and the test above is the reliable way to find out.

Is WhisperKit the same as Whisper?

WhisperKit is Argmax's framework for running OpenAI's Whisper models on Apple silicon. The model is Whisper; WhisperKit is how it runs on the Mac.

Does Parakeet support languages other than English?

Version 2 is English only. Version 3 supports 25 European languages, according to its model card. Whisper's large-v3 card lists 99.

Which model does SuprFlow use?

SuprFlow 1.0.0 uses a WhisperKit (Whisper) model that it downloads once during setup; see https://suprflow.app/requirements.

Related guides

• How to run Whisper on a Mac
• Compare SuprFlow speech models
• Offline voice to text on Mac
• Open-source Wispr Flow alternatives
• Superwhisper vs MacWhisper
• Run a fair Mac dictation comparison

Sources

• OpenAI — whisper GitHub repository and README
• Radford et al. — Robust Speech Recognition via Large-Scale Weak Supervision
• OpenAI — whisper-large-v3 model card on Hugging Face
• NVIDIA — parakeet-tdt-0.6b-v2 model card on Hugging Face
• NVIDIA — parakeet-tdt-0.6b-v3 model card on Hugging Face
• Hugging Face — Open ASR Leaderboard
• Argmax — argmax-oss-swift (WhisperKit) GitHub repository
• Handy — GitHub repository
• Spokenly — Pricing
• Superwhisper — AI models
• MacWhisper — official site
• SuprFlow — System requirements