Choosing a Mac Voice Engine: Parakeet, Whisper, Cloud, and Apple

Seven choices, not two

The Parakeet versus Whisper framing is too small for what the Mac app offers. The picker lives in Auri Voice Settings under Voice engine; each engine is a card with its name, download size, and language count, and switching takes about ten seconds.

Below the production engines, a Labs disclosure marked Experimental holds two more: Live English Dictation (Parakeet Live, ~428 MB, English-only live streaming) and High-Accuracy Multilingual (Cohere Transcribe, ~1.5 GB, 14 languages you select manually). The app's own caption applies: "Early engines for testing. Results, memory use, and language support may change."

Question one: how many languages do you speak into it

Language coverage decides more engine choices than accuracy, and it is the one question you can answer without testing.

• Only English, ever: Fast English Dictation (Parakeet TDT v2). It rates higher on accuracy and gives up every other language to get there.
• Two or more languages in a normal week: Fast Multilingual Dictation (Parakeet TDT v3), the Apple Silicon default with 25 languages. Slightly less accurate, but you stop switching engines every time you answer in another language.
• A long tail of languages: Auri Cloud covers 99 with auto-detect and needs no download, which matters when the language you need appears twice a year.

One foreign message a month does better on the English engine plus a manual fix; a life in two languages loses more time to engine switching than it gains from the accuracy difference. Try v2 for a week; if you keep opening the settings panel, move to v3.

Question two: what your Mac can carry

Local models cost disk, memory, and warmth. Those costs are invisible in a table and very visible on a fanless laptop on battery.

• Whisper Large Turbo is the high-quality local option: a ~1.5 GB download, and the app warns it needs 8 GB of RAM or more and may be unstable on some systems. Reach for it when audio is genuinely difficult: a strong accent, second-language speech, constant room noise, or vocabulary general models rarely see.
• Whisper Base is the opposite end: a ~141 MB download that runs on almost any Mac, which is why it is the Intel default. Choose it for an older machine or a nearly full SSD.
• Nemotron 3.5 ASR (~668 MB) targets the gap between speaking and seeing text. If latency breaks your writing rhythm rather than word errors, try it before you try a bigger model.

A split setup costs nothing to maintain: Whisper Large Turbo for drafting at the desk, a Parakeet engine for replies on battery.

When no download is the right answer

Two options skip the download entirely.

Auri Cloud is fast, accurate, and covers 99 languages without putting a model on your SSD. The trade: audio leaves the device, so it is the wrong choice for material your employer would not want off the machine, and when the network drops, dictation fails with an error instead of falling back to a local engine.

Apple On-Device ships with macOS 26 and costs no disk. Use it to confirm your microphone, permissions, and insertion path work before you commit disk to anything else; a downloaded engine plus custom vocabulary handles long takes and proper nouns more comfortably.

The engine is not the whole pipeline

Choosing an engine settles how audio becomes words, not what happens to those words next. Auri Voice runs a separate cleanup pass that removes filler, resolves false starts, fixes punctuation, and follows spoken formatting commands. That pass has its own provider choice in the app's AI settings, and its default is not local:

• Automatic (the default) resolves to Auri Cloud: your transcribed text, never the audio, goes to Auri's server for cleanup.
• Auri Local runs small language models (LFM2.5, Gemma) on the Mac itself; Apple Silicon only, works offline.
• Bring your own key: OpenAI, Anthropic, or Google, plus custom providers you add yourself.

Read the two settings together. A local engine with Automatic cleanup means your text goes out for polish even though your audio never did. If nothing may leave the machine, pick a local engine and Auri Local. If you already use the Auri Cloud engine, the cleanup choice is about model quality, not privacy.

This changes which errors should influence your engine choice: fillers, restarts, and missing commas get fixed for you; wrong names, wrong numbers, and homophones that read correctly do not. One take shows the split:

You said: "can you send the, um, updated deck to Anneke before Friday"

The transcript read: "can you send the, um, updated deck to Anika before Friday"

Auri typed: "Can you send the updated deck to Anika before Friday?"

Cleanup removed the filler and punctuated the sentence; it kept the misheard name because it cannot know your colleague's spelling. Fix names in Vocabulary (Settings > Dictation > Text > Vocabulary), which works with every engine: add the word, teach it by voice, or right-click a dictation in History and choose Correct with Vocabulary. The custom words guide has the playbook.

A bake-off with an email you already sent

Published error rates come from read-aloud recordings in quiet rooms; your accent, your microphone, and the twelve proper nouns you say every day are missing from them. Twenty minutes with your own material beats any table, including the one above.

1. Open a long email you sent last week, ideally with two names, a date, and a number in it.
2. Read it aloud once per candidate engine, same microphone, same room, same time of day.
3. Compare each result against the original and count two things: words you would have to fix, and seconds you waited after you stopped talking.
4. Add your recurring names and product terms to Vocabulary (it works with every engine), then rerun the leader.
5. Keep the winner for a full week of real work before you open settings again.

First fix the inputs: move the microphone to six or twelve inches instead of a laptop lid two feet away, stop switching between AirPods and the built-in mic, and speak full sentences at your normal pace. The correction count usually drops far enough that the engine question stops mattering; compare engines on the errors that survive a good setup.

FAQ