Meetings & minutes

Arrange your meeting workspace

1. Find the six-dot handle beside a card title in AI Meeting View. Drag it to a new position; the insertion line shows where it will land.
2. For click controls, open the handle menu and select an area or change the order. You can also focus the handle and use the arrow keys.
3. Click a card title to collapse or expand it. Scroll inside Live Recording to browse a long transcript.
4. Your layout is remembered on this device. Use “Restore default layout” in the meeting view to start over.

Card widths follow their content and the available workspace. Live Recording, edited text and session browsing receive more space. Empty areas collapse; smaller windows wrap cards onto another row. Moving and folding cards preserve edits in progress. Layout memory stores card positions; it does not save a session or minutes for you.

Browse and edit a meeting

1. Stop and wait for finalization, refresh Meetings & Minutes, then select a session. While recording, select the live session to view arriving segments.
2. Locate a passage using the timeline, full-text search, speaker and review filters. Long meetings are paged; search covers all segments.
3. Select a segment and review its text, time and speaker label. Save your edits before moving on. Paging or searching does not save them.

Live text correction: keep the whole original when uncertain

Requires the Windows r512 or later installer preview. Earlier previews do not include these meeting-correction checks. This section applies to supported realtime meeting processing with text correction enabled. It does not describe a shared check for every recognition route or report entry.

1. Before starting, check the recognition source language and enable live meeting text correction through the existing controls. After the correction model responds, the app checks the output language, numerical values, negation and recognized factual clues against that segment's recognition transcript.
2. Response checking has time and text-size limits. If the model request fails, times out, fails the check or cannot be checked in time, the app keeps the whole original recognition segment and separately reports that the correction was not adopted. Notices remain separate from the meeting text.
3. After selecting Stop recording, already received segments continue processing, draining and saving. Wait for finalization to end. If a task is cancelled, a derived correction arriving afterward is not published; repeatedly starting corrections will not recover that old request.
4. Compare the result with the original words and saved audio, especially names, amounts, dates and words such as “not” or “never.” When audio is available, use a passage time to listen again. Keep the original when uncertain, then edit manually.

These checks cover recognizable clues. They cannot guarantee unfamiliar Chinese names or complete semantic fidelity. Previously saved reports are not rewritten automatically. To create an updated report, review the transcript first and generate again from the appropriate report entry. Review reports and correction notices separately.

Meeting correction: protect registered names and terms

Requires Windows r513 or later. Installer previews r512 and earlier do not include this protection. It applies to supported realtime meeting text correction, not a shared check for every recognition route or LingLu report entry.

1. Open the shared text rules in Setup. Add names, brands or terms to Protected terms, for example “周琪”. Use the dictionary's “heard / write as” fields to register a correct spelling; see Dictionary import and history for batch import. No additional model setup is needed.
2. Save the rules, then start a new meeting recording. The app keeps one rule snapshot for the whole recording session. Changes saved during recording apply to a later new recording, not to segments already admitted in this session.
3. Enable realtime text correction through the existing controls. After the model replies, the app checks registered names already present in the recognition transcript and their association with negation, numbers and recognized factual clues. A similar character count is not enough. A changed name, swapped names or “not” moved to another person causes the correction to be rejected.
4. Every dictionary “write as” target acts only as a protection anchor, including targets for chat, email, code or other scopes. This check does not replace the original text, apply a scope template or send the term list as an extra model prompt. If cloud text correction is already enabled, the current segment still follows that service's existing sending behavior; names appearing in that segment remain part of its text.
5. If rules are invalid, the model fails, checking fails or times out, the app keeps the whole original recognition segment and shows a separate reason. Cancellation prevents a late derived correction from being published. Review the original words and available audio before editing manually.

The meeting does not guess recording ownership from the project currently being viewed or automatically apply that project's dictionary. Project dictionaries remain available for explicit cleanup of saved projects. Unknown names, and names already wrong in the original recognition transcript, are not guaranteed to be corrected. These checks are not a proof of complete semantic equivalence. Listen again to verify important names, amounts and negation.

This section does not enable enhanced automatic dictation delivery or change voiceprint thresholds. It does not establish speaker identity or verified real-speech Chinese quality.

Create and refine minutes

Listen to each repeated passage at its own time range before treating it as a duplicate record. Interim recognition text can still change; wait for passages to finalize before generating minutes.

Review the transcript before selecting an LLM or LingLu configuration and generating minutes. Minutes can contain summaries, sections, decisions and action items. Verify owners, dates, numbers and supporting transcript passages before saving. Do not turn unsupported model output into a meeting decision.

If live-minutes archiving or final saving fails, inspect the error and retry explicitly. A single failure does not automatically discard the session. Generation and export are separate operations.

12

1. 1Check the summary against the transcript.
2. 2Check actions, owners and dates.

Export transcripts, subtitles and minutes

Expand meeting export and choose TXT, JSON, HTML, SRT/VTT output and saved language options. Subtitle exports retain segment timing; review translation and speakers first. Minutes have their own export entry, and LingLu reports are previewed or exported from the LingLu area.

123

1. 1Choose the original or a saved translation.
2. 2Choose a subtitle or text format.
3. 3Export and choose a save location.

A session has a warning

Damaged or unreadable metadata can produce a warning on one session while other valid sessions remain available. Open the available transcript first. The warning remains until the metadata problem is resolved. Root-directory permission failures are reported explicitly; fix access and refresh.