Vocula

Vocula vs VoiceInk

VoiceInk is open source, local, and sold once — the same shape as Vocula. The easy differentiators evaporate, so this page is about the ones that survive measurement.

Buy VoiceInk instead if… you want AI cleanup of your dictation, you are on an Intel Mac or macOS 14, or you need custom vocabulary. VoiceInk is a good piece of work by someone doing the same thing we are, and on features it is ahead today.

VoiceInkVocula
Price$29, $49 or $69 once, for 1, 2 or 3 Macs. Seven-day trial.One payment, no subscription
Where recognition happensLocally, with optional cloud models if you supply an API key.On your Mac, always
PlatformsmacOS 14+, Intel and Apple SiliconmacOS 26+, Apple Silicon
Detection limited to your languages No Yes — the whole set, in your order
History encrypted at restNoYes, one sealed file per day

Where VoiceInk is better

Where Vocula is better

The hardest page on this site to write

Every lazy argument fails here. VoiceInk is GPL-3.0, so “open source” is not a differentiator. It is sold once, so “no subscription” is not one either. It transcribes on your Mac with whisper.cpp, so “local” is not one. It is a native Mac app that is notarised — which, as of August 2026, is more than we can say.

What is left has to be measured, so here it is measured.

The device ladder is priced enforcement that is not enforced

VoiceInk charges $29, $49 and $69 for one, two and three Macs. Its own source tells a different story: the licence state resolves to licensed on nothing more than a key and an activation id being present in the Keychain. There is no re-validation, ever. A revoked or refunded licence keeps working. A failure to read the Keychain resolves in the user’s favour.

We are not pointing this out to accuse anyone. We are pointing it out because we cannot enforce a device count either, and neither can any offline licence — the difference is that we are not going to charge you $40 extra for a limit that does not exist. Vocula is one price for your Macs, and the reason is written down rather than priced in.

The bug that explains why this project is careful about the VAD

VoiceInk has an open issue in which its default voice-activity detection silently discards most of a long dictation before transcription: 168 seconds of speech reduced to 103 characters, against 1,902 with the detection turned off. No error, no warning — the transcript simply arrives short and fluent.

This is the exact failure Vocula’s speech detection is built to avoid, and it is the reason the pipeline here sends one contiguous span to the model rather than stitching detected segments together. Between two segments of one sentence lives a short word, and deleting it produces a transcript that reads perfectly and is wrong.

That is not a clever marketing point. It is why the boring engineering matters in a product whose output you cannot proofread against anything.

Everything above was checked on 2026-08-21. Prices and behaviour change; if you find something here out of date, that is a bug — and here is how to check the important claim yourself.