Which model should you use in VoiceInk?

Use Parakeet V3. It is the model VoiceInk's onboarding downloads, and the maker's model guide names it the starting point for local dictation on Apple silicon ¹. It handles English and the supported European languages ². Change it only if you dictate in a language it lacks, or if you are on an Intel Mac. This page covers VoiceInk 2.x as documented on October 27, 2026. For the whole app, see our VoiceInk guide.

Parakeet V3 wins where dictation is hard: noise and accents

In a quiet room the two main local models are close. On clean read English, Whisper Large v3 Turbo had a 3.92 percent word error rate ³ against 4.48 percent for Parakeet V3 ⁴. You would rarely notice that gap.

The ranking flips under real conditions. We ran both models directly on a Mac with an M2 Pro in October 2026, outside VoiceInk, on public test sets:

On accented conversation from 13 groups, Whisper made about 40 percent more errors than Parakeet ⁹. If you speak English with an accent, or dictate the way you talk instead of reading aloud, that gap outweighs the quiet-room result. The full numbers, including long recordings, are in the VoiceInk guide.

Parakeet is also faster to use. Whisper models in VoiceInk transcribe the recording after you stop speaking, rather than streaming partial text, and larger ones take more storage and time ¹⁰.

Pick by language first, then by Mac

The language you speak narrows the choice faster than accuracy does. VoiceInk's model guide says a model being multilingual does not mean it supports every language, and tells you to check the Mode's language picker before downloading ¹¹.

Our question on which languages VoiceInk supports lists them by model, and the guide to Whisper's languages shows how Whisper's accuracy varies across them.

When another model is the better choice

You switch between languages. Make one Mode per language, each with its own model. A Mode holds the transcription model and the language together ¹⁷, so a German Mode can use one model and an English Mode another.

You dictate for minutes without stopping. In our long-recording test Whisper lost more: it left out a passage of 10 or more words in 4 of 84 readings ¹⁸, against 1 of 84 for Parakeet ¹⁹. Those engines ran with their own long-audio handling, not VoiceInk's, so treat that as a hint to test, not a rule.

You want the smallest download. VoiceInk's Whisper catalog runs from Tiny to Large v3 Turbo, with a quantized Turbo build ²⁰. Smaller Whisper models are less accurate, so only choose them if disk space is the real constraint.

You are comparing cloud options. A cloud model sends your audio to a provider and needs your own API key, and the license does not cover the charges ²¹. Everything on this page assumes local models, where audio stays on the Mac.

What to do: keep the default, then test your own voice

1. Start with Parakeet V3 and dictate your usual work for a few days.
2. If your language is not listed for it, open AI Models, then Local, and download Cohere Transcribe or a Whisper model, then select it in a Mode.
3. Compare results on your own microphone and voice. VoiceInk's guide points to History, then Analyze, for saved recording timings ²², and notes that speed and accuracy vary with your microphone, language and hardware.
4. Keep the model that makes you fix the fewest words, not the one with the best benchmark.

The same trade-offs appear in rival apps. Superwhisper offers both model families too, as the question on VoiceInk vs Superwhisper explains, and the question on VoiceInk vs Wispr Flow covers the cloud alternative. For a wider view, read the overview of dictation software and the glossary entry on on-device speech recognition.