Windows guide

How to summarize a recording locally with Ollama

Updated September 9, 2026 · ScreenBrief 0.7.8

Ollama can turn the text of a conversation into a summary on your computer. ScreenBrief lets you choose transcription and summary AI separately: audio becomes text first, then a local model prepares the notes.

What you need

Ollama requires Windows 10 22H2 or newer. See the official Windows documentation for installation and driver requirements.

• ScreenBrief on Windows 10/11 and a short test recording without private information.
• An installation of Ollama and a downloaded local model. Installing Ollama alone does not mean a model is installed.
• Disk space for models and enough memory to run them. Download size is not the same as memory requirements; choose a model your computer can handle.
• Internet access for initial installation and model downloads. Local processing does not remove network needs for updates or periodic license validation.

1. Set up local transcription

Select local Whisper transcription in ScreenBrief settings. This stage recognizes speech; an Ollama text model does not replace audio transcription. On first use, allow the required model to download.

If you select an external transcription provider, the relevant audio goes to that provider even when the later summary runs locally. Read the offline transcription guide for more detail.

2. Connect ScreenBrief to Ollama

1. Start Ollama and make sure a local text model is downloaded. A cloud-hosted model is not equivalent to a local download.
2. In ScreenBrief settings, find AI: summaries, analysis and chat, then select local Ollama.
3. For Ollama on the same computer, use the default connection. The standard local address is http://127.0.0.1:11434. Use “Ollama runs at a different address?” only for a custom address; a remote server receives your text.
4. Save the connection and use “Test connection”. If the model list is empty, download a local model in Ollama and test the connection again.
5. Choose an available model. The local Ollama connection does not require a paid provider API key.

3. Start with a short recording

Screen analysis is an additional stage. Before enabling it, check the selected provider and the model’s image support. Choosing local transcription alone does not establish that every stage runs locally.

1. Import a short recording containing one clear decision and one action item. Disable screen analysis for the first test so you can check audio, text and summaries alone.
2. After processing, read the transcript. Check names, numbers and sentences containing “not”: recognition errors can change the meaning of the notes.
3. Review the summary, decisions and tasks. Compare them with the relevant transcript passage and playback.
4. Try a longer recording after the short test succeeds. Speed and quality depend on the model, text length and computer.

Check the facts behind fluent text

Try this example: Anna says “We are choosing the blue version of the page”, Jan adds “I will prepare the mockup by Friday”, and Anna finishes with “We will agree the budget later”. A summary should preserve the color decision and Jan’s task. The budget remains an open issue without an agreed owner or deadline.

If the model invents an owner or deadline, or turns a suggestion into a decision, return to the source. You can also ask the chat which passage supports a conclusion and check its reference. Do not accept an unverified task just because it sounds plausible. In a ScreenBrief 0.7.8 test with local qwen3-vl:4b, a 20-second recording was transcribed correctly, but the summary invented context and the decision list omitted the color choice. Chat correctly identified the blue version and cited 0:02; clicking the citation moved the player to that statement. This is one short test, not a guarantee of model quality.

Troubleshooting

• No connection: check that Ollama is running in the background and the saved address points to the right computer. Then repeat the connection test.
• Connected but no models: download a local model in Ollama. An account with access to a cloud model does not replace a local model.
• Very slow generation or insufficient memory: try a smaller model and a shorter recording. There is no single processing time for every computer.
• Missing decisions in the summary: first check the transcript, then compare another model using the same short example. A connection test checks availability, not note quality.

Cost and the limits of local processing

ScreenBrief includes 5 transcriptions at no charge. Further transcriptions require Pro for one installation. A local model does not incur summary API charges, but uses your computer’s resources.

Local Whisper and local Ollama let you process audio and text on your computer. Selecting an external provider or remote address changes where data goes. See the Ollama FAQ for local versus cloud models, and the local meeting notes guide for ScreenBrief configuration.

Details of how data is stored and transferred: privacy policy.