Free during the beta

Transcribe meetings and dictate. On your computer, nowhere else.

Onsite records Teams, Zoom and Google Meet calls on the Mac and writes the transcript there, with speakers and timestamps. It transcribes audio files you already have the same way, and takes dictation into any program. German and English, no account, no cloud.

One window with the navigation on the left. Meetings have titles, tags and full-text search, with the transcript beside them.
One window with the navigation on the left. Meetings have titles, tags and full-text search, with the transcript beside them.

Record, transcribe, dictate. On the machine.

Nothing leaves the machine

The recording and the transcript are files in a folder on the laptop. Your IT department can confirm that with Wireshark in ten minutes.

Title and summary by themselves

After the meeting a language model on the Mac writes a short title, a summary and the key points, in about two and a half minutes for a 50-minute call. The summary sits in the meeting list, so you need not open the transcript. The model is a one-time 6.2 GB download that you start yourself.

Speakers by name

Every paragraph says who spoke and when. NVIDIA's Nemotron speech model tells the voices apart, and you enter the names once. Any program opens the transcript, and Onsite exports it to Word, PDF or text.

Asks before it records

Onsite notices when a call starts in Teams, Zoom, Meet or another meeting app and asks you. A browser counts only with a call in its window title or sound going both ways, and other programs do not prompt. No bot joins, and without your yes nothing is saved. While it asks, it listens in memory so the start of the call is kept.

German and English, on the Neural Engine

Both languages, in the same meeting if need be: a German call stays German, and English stays where it was really spoken. On German FLEURS, a public test of 18,577 words, 5.5% of words come out wrong, against 8.6% for the recogniser Onsite 1.1 used. Parakeet Ultra runs on the Neural Engine, which leaves the processor and the graphics to the call, and the transcript is ready a minute or two after the call ends. The program itself speaks German too.

Dictation into any program

Press ⌃⌥D, speak, press it again. The text lands where the cursor is: Word, Outlook, the ticket system, about 0.4 seconds after the second press.

The recordings you already have, too.

Drop voice memos, phone recordings or interviews anywhere on the window, or choose them on the Transcribe page. WAV, AIFF, MP3, M4A, FLAC and Ogg are transcribed one after another and join your meetings, with speakers, a title and search.

The Transcribe page with its drop zone and queue.

You do nothing during the meeting.

When a call starts, Onsite asks whether to record. It captures your microphone and the other side separately, and follows you when you switch headsets or speakers mid-call. If a long stretch is lost anyway, the transcript says so. A few minutes after the call ends, the transcript is in Documents/Meetings, titled by what the call was about. If there was no real conversation, Onsite discards the recording, tells you, and keeps it under Recently discarded for seven days.

One window.

The interface is in English and German.

Questions

Do we need a data-processing agreement?

No. The vendor never receives any data, so there is no processor to contract with.

How does IT verify that nothing leaves the machine?

With Wireshark, in ten minutes. Onsite makes three kinds of connection: the model download, the update check and sending a feedback report. It starts the model download, automatically whenever a model is missing at launch (on first start, after an update that adds one, or to finish a failed download); everything else only when you click a button. The steps are on the IT review page.

Can an employer use it to monitor staff?

No. There is no admin view, no central store and no analysis of staff. It keeps a recording only after the user has agreed.

Onsite 1.3 for the Mac.

After the 14-day trial you ask for a free key by email, as the pricing page describes. On first launch Onsite downloads the speech models on its own, 845 MB, without a click. The installer is not signed yet, so macOS warns once on first launch.

macOS

A Mac with Apple Silicon, 16 GB of memory or more, and macOS 14.6 or later. Macs with less memory and Intel Macs are not supported.

At the Gatekeeper message, choose “Open Anyway” under “Privacy & Security”.

Download for macOS

Rolling it out to a team: the IT review page. Every version with its checksum: the releases page.

Write to us.

The form opens your mail program. This website stores nothing you type.