Just Transcribe, in plain numbers.

In one paragraph

Just Transcribe is a voice to text app for iPhone, iPad, Mac and Apple Watch, made by a one person studio in the Netherlands. It records audio, or imports a file you already have, and turns it into text using speech models that run on the device itself: 21 in total, from five engines. Nine come from the Whisper family (six released by OpenAI, three distilled or optimised variants made by others), and the rest are NVIDIA's Parakeet v3 and Canary 1B v2, two Qwen3-ASR models for East and Southeast Asian languages, and eight builds of Meta's Omnilingual ASR, which covers more than 1,600 languages. Your audio and your transcripts never leave the phone, because there is no server to send them to and no account to sign into. The download is free and two models are free forever; a single purchase of EUR 9.99 unlocks the rest, once, with no subscription.

Price, in full

There is one purchase in the app and you never have to make it twice; the Mac app has a separate one of its own. The free tier is a working transcription app rather than a countdown to a paywall: you can record, import and transcribe with two real models for as long as you like.

The EUR 9.99 unlock opens: the 19 speech models beyond Tiny and Base, follow along highlighting, speaker identification and the voice library, transcript editing with find and replace, on-device translation, on-device summaries, Add punctuation, audio enhancement, Advanced Privacy, export to SRT, VTT, PDF, text and Markdown, the Apple Watch app, CarPlay recording, and sending recordings to the Mac app. The model advisor and Siri access are free. On the Mac, the EUR 19.99 unlock opens the 19 bigger speech models, follow along, speakers, translation, Add punctuation and the link with the iPhone.

Privacy, specifically

"Private" is a word every app uses, so here is what it means here, item by item. The full privacy policy has the legal version.

One honest footnote, because it is the sort of thing worth being straight about: this website uses Google Analytics, and only after you click accept on the consent banner. The app uses no analytics at all. Those are two different things.

Advanced Privacy, the mode that erases itself

A separate mode inside the app, opened with the lock in the top left corner of the main screen. It runs one recording from start to finish and then wipes it: the recording never enters your library, the audio is erased the moment the transcript has been read out of it, and whatever is left goes when you close the screen. The app already does all of its work on the device in every mode; this is the mode where nothing is kept at all, and where the claim can be checked with the radios off. There is a full page about it.

The 21 speech models, and who is in charge of them

You decide which models exist on your phone. Nothing is downloaded until you ask for it, you can keep several side by side, switch per recording, and delete any of them again from the model manager. That is the whole point: the engine is a file you own, not a service you rent.

There are 21 of them, from five engines, and the model manager opens with four questions (which language, what you are recording, speed or accuracy, follow along or just text) that point at the one that fits. Speed is an estimate for a recent iPhone, in seconds of processing per minute of audio, so lower is faster. Quality is the app's own indicator on a scale of five.

Six models carry the top rating of 5 out of 5: Large v3, Large v3 Turbo, Canary 1B v2, Qwen3 ASR 1.7B and both Omnilingual 7B builds. Large v3 Turbo is the fastest of them. The two free models are honest about their limits, and the app says so on screen: Tiny and Base do well with clear speech and one or two people, and struggle with strong accents, background noise or a full meeting table.

Three things differ per engine. Follow along and speaker labels need word timings, which Whisper, Parakeet and Omnilingual deliver and Canary and Qwen3-ASR do not: those two give text only, and want the language picked by hand. Omnilingual needs no language at all and writes what it hears in the right script, without capitals or punctuation, which Add punctuation puts back. And speech straight to English works on the multilingual Whisper models only, not on the two Distil ones, Parakeet, Canary, Qwen3-ASR or Omnilingual. The app hides an option a model cannot do rather than quietly producing nonsense.

Every model and framework is listed with its maker and licence in Settings, under Models & licences.

Languages

Transcription: 59 languages to choose from, plus automatic detection. Version 1.36 added nine: Cantonese, Filipino, Indonesian, Japanese, Korean, Malay, Persian, Thai and Vietnamese. Pinning the exact language gives sharper results than letting the model guess, and the picker only shows what the chosen model can do.

Albanian, Arabic, Armenian, Azerbaijani, Basque, Belarusian, Bengali, Bosnian, Breton, Bulgarian, Cantonese, Catalan, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Faroese, Filipino, Finnish, French, Galician, Georgian, German, Greek, Hindi, Hungarian, Icelandic, Indonesian, Italian, Japanese, Korean, Latvian, Lithuanian, Luxembourgish, Macedonian, Malay, Maltese, Norwegian, Norwegian Nynorsk, Occitan, Persian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swedish, Thai, Turkish, Ukrainian, Urdu, Vietnamese and Welsh.

Beyond the list: Omnilingual ASR recognises more than 1,600 languages by itself, hundreds of which no other model in the app covers. There is nothing to pick. Coverage of the rarest ones is uneven, so try a short recording first.

What is in it

Requirements

What it deliberately does not do

Half of choosing an app is knowing where it stops. None of the following is on a roadmap being coy about it; these are choices.

• ×No cloud sync between devices. Nothing syncs, because nothing is uploaded. Your iPad does not see your iPhone's recordings.
• ×No team or shared workspace. This is a single person's app on a single person's device.
• ×No live captions. Transcription happens after the recording, not word by word while someone is still talking.
• ×No meeting bot. Nothing joins your Zoom or Teams call.
• ×Not in a browser or on Windows. It runs on iPhone, iPad, Apple Watch, Android and Macs with macOS 26 or later, and there is no web or Windows version.
• ×No subscription, account or data collection. Which is the point, but it is still a limit: there is no dashboard, no login and no way to recover anything from a server, because there is no server.

Who it suits

It fits particularly well when the content of a recording is confidential and sending it to a cloud transcription service would be a problem in itself.

• ✓Journalists. Source protection is not compatible with uploading the interview to somebody else's server.
• ✓Healthcare and therapy. Session notes that legally should not leave the room, and certainly not the country.
• ✓Law and HR. Privileged and sensitive conversations, without a data processing agreement for every voice memo.
• ✓Researchers and students. Interviews under an ethics approval that says the audio stays local.
• ✓Councils, boards and works councils. Minutes from a meeting that is not public yet.
• ✓Anyone on a plane, a train or a bad connection. Offline is not a fallback here, it is the normal mode.
• ✓Anyone tired of renting software by the month. One purchase, kept for good.

Straight answers