Experimental

AI tools: summarize, minutes, translate

A transcript is raw material. The AI actions turn one into something you'd actually send — a summary, a set of meeting minutes, or the same conversation in another language — using an AI engine you choose and configure yourself.

Last updated 2026-08-31

The three actions

Open a transcript in the History window and you'll find three actions:

Summarize

A condensed account of what the transcript covers, for when you need the gist without re-reading an hour of talk.

Meeting Minutes

Structured notes: what was discussed, what was decided, what someone agreed to do. Best on a transcript that has speaker labels.

Translate

The transcript rendered into another language, keeping the shape of the original.

Results stream into a sheet as they're generated, so you watch the text arrive rather than staring at a spinner. When it's done, the result is saved onto that history item — reopen the transcript tomorrow and the summary is still there, no re-run needed.

These are marked experimental, and we mean it. Quality depends heavily on which model you point them at. A small local model produces a serviceable summary; a large one produces something closer to what a person would write. Read what comes out before you send it anywhere.

You bring the engine

LymeScribe does not ship an AI model for this and does not run one for you. Instead you configure one or more engine profiles, and every action runs against the profile you pick. Any endpoint that speaks the OpenAI chat-completions API works, which in practice means two kinds of setup:

Setup	What it looks like
A local server	Something like LM Studio running on this machine or another computer on your network, serving a model you downloaded. Nothing leaves your network. Usually no API key needed.
A cloud provider	A commercial API endpoint. Better results from bigger models, at the cost of sending the transcript to that provider. Requires an API key.

Each profile holds a display name, the base URL of the endpoint, the model identifier to request, a context budget (how much text the model can take at once), and — for cloud endpoints — an API key. Keys are stored in your operating system's secure credential store, not written into the app's settings file.

You can keep several profiles side by side (a fast local model for quick summaries, a bigger one for minutes), set a default, and switch per action.

Setting one up: the AI tools setup walkthrough takes you through it step by step, for both the local-server and cloud-provider routes. If you want better output from whatever engine you've got, see prompt recipes.

Where your transcript goes

This is the part to be clear about, because it's the one place in LymeScribe where data can leave your machine.

Transcription itself is always local — that never changes. The AI actions send the transcript text to whatever endpoint your chosen profile names. Point that profile at a local server and the transcript never leaves your machine or your network. Point it at a cloud provider and the transcript goes to that provider, under their terms, exactly as if you'd pasted it into their web app yourself.

Nothing leaves the machine unless you configure a profile that points off it. There is no default cloud engine, no LymeScribe-operated service behind these actions, and no telemetry carrying transcript content anywhere. If you never set up a profile, the actions simply have nothing to run against.

Handling anything confidential — client matters, medical notes, privileged conversation — is the case for a local engine, and it's why the local route is documented first in the setup guide.

Long transcripts

Every model has a limit on how much text it can consider at once, which is what the profile's context budget records. A two-hour meeting will exceed most of them.

LymeScribe handles that for you. If a transcript fits, it goes through in one pass. If it doesn't, the app splits it into chunks and works through them — for Summarize and Meeting Minutes it processes each chunk and then combines the pieces into one result; for Translate it works through the chunks in order so the output reads continuously. You don't configure any of this; you just notice that long transcripts take longer.

Setting an honest context budget on the profile matters here — too high and requests get rejected by the engine, too low and long transcripts get chopped more finely than they need to be. When you pick a model from a local server, LymeScribe fills the budget in from what the server reports.

Running an action

Set up at least one engine profile in Settings ▸ AI (see the setup walkthrough).
Open the History window and select the transcript you want to work on.
Choose Summarize, Meeting Minutes, or Translate.
Watch the result stream in, then copy it out or leave it saved on the item.

When an action fails

Can't reach the engine. For a local server, confirm it's actually running and that the base URL and port match. For a cloud endpoint, check the URL and your network.
Rejected or unauthorized. Usually a missing or stale API key — re-enter it on the profile.
Unknown model. The model identifier has to match exactly what the endpoint serves. Local servers list theirs; pick from that list rather than typing from memory.
The request is too large. Lower the profile's context budget so chunking kicks in sooner.
The output is thin or wanders. That's the model, not the app. Try a larger one, and see prompt recipes.

You don't have to use these at all

The AI actions are a convenience, not a gate. Export the transcript and paste it into whatever AI you already use — that's the workflow described in from meeting to meeting notes, and it works just as well. The in-app actions exist so you don't have to leave the app, and so people who want a fully local pipeline can have one.