Whisper or Parakeet: the two engines, and how to choose

Vocula is a $29 one-time dictation app for Apple Silicon Macs on macOS 26 or later. It ships two recognition engines and you pick which one runs. This page is the whole difference between them — including the one Parakeet is worse at, which is the reason it is not the default.

Download figures are byte counts of the two assets on the project’s own release: 1 624 555 275 and 466 638 996. Add 885 KB for the voice-activity detector either way. Only the engine you select is downloaded.

Where the speed figure comes from, and what it is not

Five recorded fixtures, utterances of 1.4 to 2.5 seconds, run back to back on one machine — an M5 Pro. Parakeet returned the text in 57–76 ms, Whisper in 217–230 ms. Both engines transcribed all five identically.

Every word of the qualifier matters: short phrases, one Mac, five files. It is not a claim about your Mac, it is not a claim about long dictation, and it is emphatically not an accuracy comparison — five clean fixtures agreeing tells you they agreed on five clean fixtures. There is no multiple on this page for the same reason there is none anywhere else here: the two ranges are in the sentence above and you can do the division yourself, including at both ends of it.

We have not measured word error rate for either engine in any language. NVIDIA advertises 2.6 % for Parakeet v3; that is an average across its 25 languages and it is not the figure for yours. Nobody here has measured a Cyrillic error rate, so there is no number on this page for it.

The reason Parakeet is not the default

Parakeet covers 25 European languages. It cannot transcribe Japanese, Chinese, Arabic or Hindi at all.

A default may be slower, or occasionally wrong. It must not be incapable: someone whose language is simply absent from the screen has no way to discover that the other engine would have served them. So Whisper stays the default, and a test in the app repository pins that decision so it cannot be reversed by accident.

The failure mode worth knowing before you switch

Parakeet holds to one alphabet per phrase. An English technical term inside a Cyrillic or Greek sentence may come back transliterated, or dropped, where Whisper keeps it. Say “сделай pull request” and the two words in Latin script are the ones at risk.

This is not a setting we have failed to add. It is how the model is built. Parakeet has 8,192 tokens for all 25 languages, of which 76 % carry Latin letters and 14 % Cyrillic; Whisper has 51,866. No word is a single token — pull, request, GitHub, commit, merge, branch are each assembled from pieces, and every piece is a separate decision that leans on the one before it. Measured on the vocabulary file the app ships.

Pinning a language makes this more likely rather than less, because the filter then removes the foreign alphabet from the candidates entirely.

FluidAudio’s own issue #512 describes wrong-alphabet tokens, and we reproduced the behaviour in ordinary use — two dictations of the same term came out differently on different days. That is the honest description: a bias, not a rule.

What the language switch does on each engine

Vocula’s wedge is that recognition is restricted to the languages you chose rather than pinned to one or guessed across a hundred. That describes Whisper precisely, and it describes Parakeet only in part.

• On Whisper, detection is an argmax over your list. A language you did not tick cannot be returned. This is the claim no competitor answers.
• On Parakeet, with automatic detection on, the list does not constrain anything: the model identifies the language inside the same pass that transcribes it, across all 25 it knows. We verified this the unflattering way — a Polish sentence dictated with Polish not ticked came back as correct Polish. Turning automatic detection off and pinning one language does reach the engine, where it restricts the output to that language’s alphabet.

Half of a product’s pages saying something that is true of the other half is how a site quietly stops being checkable. So it is written here instead.

Both engines are still local, and still verifiable

Nothing about the privacy position changes with either. Both models are downloaded once, from Vocula’s own GitHub release rather than a third-party model host, and pinned by SHA-256. Both run in a process you can watch with nettop, and neither opens a socket while it transcribes.

One detail belongs here because it is specific to Parakeet: FluidAudio’s loader will fetch a missing file from HuggingFace by itself, and Vocula refuses to load the engine unless every bundle is already on disk. How that was found, and how to check it.

What is still unknown

• Whether pinning a language helps on live bilingual speech. Unproven. On clean synthetic fixtures the model separates Polish from Russian perfectly, so the hint changed nothing in ten comparisons. On a real speaker it is expected to matter; nobody has measured it.
• Cyrillic word error rate for Parakeet. Unmeasured, and the advertised 2.6 % average does not answer it.

When those are measured they go on this page, naming the Mac and the conditions. Until then their absence is the honest version.

Common questions

Download Vocula for macOS — $29 once, no subscription, no account · What it costs · Every Mac dictation app, and the engine each one ships