How to summarize a long recording locally

The short answer

Summarizing is two steps: transcribe the audio, then have a language model condense the text. Both steps can run on a modern computer. Use a local transcription model, then a local language model, and nothing leaves your machine.

Step 1: get a good transcript

A summary can only be as good as the transcript under it. Transcribe with a model that suits the audio, add names and technical terms if your tool supports it, and make sure speaker labels are there: "we agreed" is only useful if you know who "we" were. See transcribing offline on a Mac for the details.

Step 2: summarize without the cloud

Local language models are smaller than the big cloud ones, so they work best with a clear job: a summary, a list of chapters, action items. Very long transcripts may not fit in a small model's working memory at once, so good tools split the recording into sections, summarize each, and combine them.

The trade-offs are speed and polish. A local summary may be plainer than a cloud one, and a bigger model needs more memory. In return you get no upload, no per-token billing and no account.

What to ask for

"Summarize this" gets you a generic paragraph. Asking for a specific shape usually gets more useful output from any model:

• Chapters with timestamps, so you can jump to the part you need.
• Decisions and action items for meetings, with who owns each one.
• Key arguments and evidence for lectures and talks.
• Direct quotes for interviews, which you can then check word for word against the audio.

How to do it in MediaFind

Summaries, chapters and Ask are in the free plan and run on an on-device language model.

1. Add the recording, or the folder it's in. MediaFind transcribes it on your computer.
2. Open the file to see its summary and its auto-segmented chapters, so you can jump straight to the part you need.
3. Ask follow-up questions, e.g. what were the three objections the client raised? Answers cite the moments they came from, so you can click through and check.
4. For a meeting, Meetings mode (Pro) turns the recording into notes with key points, action items and decisions, exportable to Word or PDF.
5. Export the transcript as Word, PDF or text if you want to edit the summary yourself.

How to check a summary

Any language model, local or cloud, can leave things out or state them more firmly than the speaker did. Before you forward a summary: check every number and name against the transcript, click through the cited moments for anything that matters, and skim the chapters for a topic the summary skipped.

The first time you use Ask, a small model downloads once; larger ones are a one-time download too. After that it works offline. More on the approach in chapters and highlights and running a local LLM.

Frequently asked questions

Can I summarize a recording without uploading it?

Yes. Transcribe it with a local speech model, then summarize with a local language model. MediaFind does both on your computer.

How long a recording can be summarized locally?

Long ones work when the tool summarizes section by section. MediaFind also splits long files into chapters, so you can read a summary per part.

Are local summaries accurate?

They are usually good at the gist and can miss details or nuance. Check names and numbers against the transcript, using the cited moments.

Do summaries cost extra in MediaFind?

No. Summaries, chapters and Ask are part of the free plan and run on your device.