On-device meeting transcription software. Softspoken.

On-device, offline and local: the same thing?

People search for this in three ways: on-device, offline and local meeting transcription. The words overlap, and they answer two different questions.

On-device meeting transcription says where the speech becomes text: on your own computer, by a speech model that runs there. Local meeting transcription means the same thing. Where the transcript is kept afterwards is a separate question, so ask about both.

Offline meeting transcription says what happens with no connection. On-device transcription makes it possible, because a model on your machine has no need of the internet to do its work. Whether a given product keeps working offline is still a separate thing to ask. Some notetakers can record with no connection and transcribe once it returns, which is offline recording with the transcription still done on a server.

On-device transcription is still uncommon. Most meeting notetakers send the audio to a server for that step. Softspoken does it on your computer, and it is not alone in that, so the rest of this guide is about how to weigh any product that says it does.

How on-device meeting transcription works

There are three steps. With on-device transcription the first two stay on your computer, and the third is the one to ask about. Softspoken is the example here.

Capture

Softspoken does not replace your meeting app. You hold the meeting in Zoom, Teams, Google Meet or whatever you use today, exactly as you do now, and Softspoken runs beside it on your computer.

It records two things as separate channels: your microphone, and the sound your computer is playing, which is the other side of the call. Nothing joins the meeting, so there is no bot in the participant list, and there is nothing to install in the meeting app or to get approved. Recording from the computer like this is common now, and several notetakers do it. For a meeting in a room there is no call to capture, so the laptop's microphone records the conversation.

Transcription

A speech model on your machine turns the audio into text, using your own processor. This is the step most notetakers hand to a server, and it decides whether a recording of your meeting leaves your computer. With Softspoken the audio is never uploaded. The transcript is an ordinary file on your own disk. By default the recording is deleted when your notes are done, and you can choose to keep it.

Notes and summaries

A transcript is the full record. A summary with action items is a further step, written by a larger AI model. On-device transcription says nothing about where that step runs, so it has its own question in the checklist below. In Softspoken it is called the AI layer, and it is optional: leave it off and you keep the full transcript.

The diagram on How it works shows the same steps, and what is sent under each AI choice.

What it asks of your computer

Running the speech model yourself asks a few things of you and of your computer. It is better to know them before you choose.

What has to be downloaded? The speech model for meetings is a one-time download. It arrives the first time you need it, and that step needs a connection. Dictation works from the first launch, with nothing more to fetch.
Which computers can run it? An Apple silicon Mac running macOS 15 or newer, or a 64-bit Windows computer. The requirements answer has the versions.

What you get in return: once the model is on your machine, recording, transcription and dictation keep working with no connection at all. Summaries, sign-in and updates wait until you have one.

Names, jargon and who said what

Any speech model can mishear a word it has never met, such as a client's surname or an acronym from your field. What matters is whether you can teach it.

Softspoken has a Dictionary for this. You add the names and terms of your work, and how they tend to be misheard. Softspoken fixes their spelling in a transcript, and gives them to the AI layer so that a summary has them right. When you spot a mistake in a transcript, you can select it and add the right term from there. The Dictionary is kept on your computer.

Who said what depends on the meeting. On a call, the two channels settle it: your microphone is you and the call's audio is everyone else, so your words are kept apart from the other side's. The other side stays together, as one. In a room there is a single microphone, so Softspoken tells the voices apart by how they sound, and you give each speaker a name. It can learn your own voice, so that your lines are marked as yours. This part is good, not perfect, so read the notes before you send them on.

A name you give a speaker also gives that person a page of their own, which lists your meetings together.

What to check before you choose

These five questions suit any product that says it transcribes on the device, ours included. A vendor's privacy or security page should answer each one plainly. Our own answers follow each question.

Where is the speech transcribed? The answer to look for is on your own computer, for every meeting. Softspoken transcribes there and never uploads the audio.
Does it keep working with no connection? Recording offline and transcribing offline are different things, so ask which one is meant. Softspoken does both, once its meeting model has been downloaded.
Does it cover a meeting in a room? Softspoken does, through the laptop's microphone. Several other notetakers cover in-person meetings too.
Is it the same on Mac and on Windows? Softspoken ships the same features on both. Several others have desktop apps for both as well.
What leaves your computer after the transcript is written? Ask what is sent to write a summary, where it is processed, and where the transcript is stored. With Softspoken the transcript is stored on your own disk. The AI layer is optional, and when you use it, text is sent for that step and audio never is. Softspoken AI processes the text in Australia and returns it without keeping it. On a paid plan the AI layer can run on your own Anthropic or AWS account instead, and then the text goes to that provider directly, not through us. If you connect an AI tool of your own to your meeting library, which a paid plan allows, the text you ask it about goes to that tool. You can also keep one client's folder off AI: its meetings are still recorded and transcribed on your computer, and are never sent to any AI model.

On the third and fourth questions the well-known products are often level with us. They differ on the first two: most send the audio to a server to be transcribed, so they need a connection to do it. Our guide to private AI notetakers sets six of them beside Softspoken, row by row, as does the comparison on our home page.

What this guide says about other products was read on their own websites on 12 October 2026.

Where Softspoken fits

Softspoken is for people whose meetings are confidential and who want the notes written for them. It is an addition to the tools you already use, and it takes the place of none of them. Because it listens on your computer, it works with every meeting app and needs no integration with any.

It transcribes on your computer, on Mac and on Windows, and you can file meetings in a folder for each client or project.

The app that records, transcribes and searches your meetings costs nothing, and needs no account or card: the Free plan is A$0 forever. The paid plans are about the AI layer. Pro is A$19 a month, and Perpetual is A$299, paid once, with AI running on your own keys. The pricing page sets out each plan, the free plan's AI allowance included.

It is in closed beta for now, and the waitlist is how invitations go out.

On-device meeting transcription software: how it works and what to check