The best local speech-to-text models for a Mac, compared from their model cards

For dictation on a Mac without the cloud, four model families come up again and again: OpenAI's Whisper Large v3 Turbo, NVIDIA's Parakeet TDT v3 and Nemotron ASR, and Cohere Transcribe. Whisper covers by far the most languages. Parakeet and Nemotron are smaller and built for speed. Cohere Transcribe covers 14 languages, and superwhisper calls it its most accurate local model.

I make YapToText, which runs Whisper, so I have a stake in this. The facts below are from each model's own card, read on October 9, 2026, and I do not publish an accuracy ranking of my own.

Download YapToText on the Mac App StoreThis app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.

Side by side

Model cards on Hugging Face, read October 9, 2026. Cohere Transcribe's card needs a sign-in, so its row is from superwhisper's voice models page.

Whisper Large v3 TurboParakeet TDT 0.6B v3Nemotron ASR StreamingCohere Transcribe
Made byOpenAINVIDIANVIDIACohere
Size809 million parameters600 million600 million1.3 GB download, per superwhisper
LanguagesAbout 10025 European languagesEnglish, or 40 language-locales in the newer Nemotron 3.5 version14, per superwhisper
Built forAccuracy across languagesFast transcription with timestampsLow-latency streamingAccuracy, per superwhisper
LicenseMITCC-BY-4.0NVIDIA Open Model LicenseSee its gated model card

What each one is good at

  • Whisper Large v3 Turbo. A pruned Whisper Large v3, with its decoder cut from 32 layers to 4, which its card says makes it much faster at a small cost in accuracy. The choice when you dictate in many languages.
  • Parakeet TDT 0.6B v3. Punctuation, capitalization and word-level timestamps, in 25 European languages. A strong pick for English and European dictation where speed matters.
  • Nemotron ASR Streaming. Designed to stream, with chunk sizes from 80 milliseconds up, so words can appear as you speak. The original is English only. Nemotron 3.5, released June 4, 2026, covers 40 language-locales.
  • Cohere Transcribe. superwhisper's docs call it the most accurate local model in its app, with an 8.4 percent word error rate on an Apple M4 by their measurement, in 14 languages.

There is no single best model: the right one depends on your language first and your patience second.

Which Mac apps offer which

  • Whisper. YapToText, superwhisper, VoiceInk, MacWhisper, Vowen and FluidVoice.
  • Parakeet. superwhisper on Pro, VoiceInk, Vowen, FluidVoice and MacWhisper.
  • Nemotron. FluidVoice and Vowen.
  • Cohere Transcribe. superwhisper on Pro, and FluidVoice.

From each app's own pages, read October 9, 2026. The Parakeet vs Whisper page goes deeper on those two.

How to choose

  • You dictate outside Europe's languages. Whisper, which covers about 100.
  • You want words on screen as you talk. A streaming model such as Nemotron.
  • You dictate in English or a European language and want speed. Parakeet v3.
  • You have an older or smaller Mac. A smaller Whisper model, or Parakeet, will load faster than the largest models.

Whatever you pick, test it on your own voice and microphone for a day before you settle. Published word error rates come from benchmark audio, not from you.

What YapToText uses

YapToText runs Whisper on the Mac, and its App Store build ships with a speech model already inside, so there is nothing to download first. It also runs a separate cleanup model on the Mac. It does not offer Parakeet, Nemotron or Cohere Transcribe, and if you want those, the apps listed above are where to look.

Download YapToText on the Mac App StoreThis app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.

The Whisper Large v3 Turbo page and the local Whisper guide cover running Whisper on a Mac.

The short version

  • Most languages. Whisper Large v3 Turbo, about 100, MIT license.
  • Fast European dictation. Parakeet TDT 0.6B v3, 25 languages.
  • Streaming. Nemotron ASR, English or 40 language-locales in version 3.5.
  • Accuracy in 14 languages. Cohere Transcribe, by superwhisper's account.

If a model card changes, or I have missed an app, open an issue on GitHub. Every one gets read.

Download YapToText on the Mac App StoreThis app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.

About the author. Ryleigh Newman is a developer whose hands make typing difficult and who dictates every day with YapToText, the app this site describes. Facts about other apps come from their own sites, with the date they were read. More about the project.

Sources

Questions

What is the best local speech-to-text model?

It depends on your language. Whisper Large v3 Turbo covers about 100 languages. Parakeet TDT v3 covers 25 European languages and is built for speed. Nemotron ASR is built for streaming.

Parakeet v3 vs Whisper Large v3 Turbo: which is better?

Parakeet v3 is smaller, at 600 million parameters against 809 million, and covers 25 European languages. Whisper Turbo covers about 100. Test both on your own voice.

What is Nemotron ASR?

An NVIDIA streaming speech model with 600 million parameters. The original is English only, and Nemotron 3.5, released June 4, 2026, covers 40 language-locales.

What is Cohere Transcribe?

A speech model from Cohere. superwhisper's docs call it their most accurate local model, in 14 languages, at 1.3 GB.

Which model does YapToText use?

Whisper, running on the Mac. The App Store build includes a speech model, so it works from the first launch.