Meeting Minutes from Your Transcript

Most meetings leave two kinds of trace. One is the memory of the people in the room, which fades by Friday. The other is the minutes, which are meant to outlast it.

What minutes are for

The word gives the game away: minutes means notes, not sixtieths of an hour, and these notes are the official written record of what a group did. Robert’s Rules of Order, the most widely used manual of parliamentary procedure in the United States, gives the job of keeping them to the secretary, and the advice that follows from it is blunt: minutes should hold mainly a record of what was done at the meeting, not what was said by the members. A board’s minutes can be a legal document kept on file for years. A committee may keep only informal notes. Either way, the reader is someone who was not there, or who was there and has since forgotten.

That reader has narrow questions. What did we decide? Who said they would do what, and by when? Was the report accepted or sent back? A page that answers those questions in a minute of reading has done its job. A page that makes the reader dig has not, however faithful it is.

A transcript is not minutes

When a meeting has been recorded and captioned, it is tempting to think the matter is settled. Everything is there, so nothing can be lost. The long-standing guidance on minutes points the other way. For most groups a verbatim report is typically not useful, and a summary of the discussion is neither necessary nor appropriate unless the group’s own rules ask for one. Minutes are a record of the meeting, not a transcript of it. An hour of talk fills pages, and the decisions taken in that hour may fit on an index card.

The trouble is not length alone. Discussion and decision look the same on the page. Five people float an idea, two object, one restates it with a change, and the chair says that will do. Only the last line is the decision.

This is where a written summary can go wrong in two directions at once. It can drop a decision that was made quietly, in half a sentence near the end, because it did not sound important. It can also invent one, by turning a discussion that trailed off into a conclusion nobody reached. Both errors are hard to catch later, because the summary reads smoothly and the words that would contradict it are buried in a recording nobody wants to replay. The safer habit is to keep each decision exactly as it was spoken, attached to the second it was spoken, so anyone in doubt can jump to that moment and listen for themselves.

A caption file already divides a meeting into short timed pieces, called cues. Marking one as a decision costs a keystroke. Its words stay untouched, and so does its time.

Decisions, actions and who owns them

An action item, in the plain management sense, is a documented task that one person can handle, usually created when a discussion discovers that something needs doing. It is assigned to someone, typically a member of the group, who is then expected to do it and report back. A usual record names the person, describes the task and notes when it was raised. The two fields that matter most are also the two most often left blank: the owner and the due date. “We should check the hive suits” is not an action. “Member A checks the hive suits by 29 May” is.

Owners go missing because tasks are agreed in the passive voice. Somebody says a thing needs doing, everybody nods, and the minutes faithfully record a task that belongs to no one.

Dates go wrong more quietly. A date typed as 30 February looks like a date until a calendar refuses it, and a due date earlier than the meeting is usually a slip for the following month. Writing dates in the ISO 8601 form, year then month then day as YYYY-MM-DD, removes the doubt over whether 04/03 means April or March and makes checking mechanical. Is the month between 1 and 12? Does the day exist in that month, remembering that February gets a 29th only in years divisible by 4, except century years not divisible by 400? Is the date on or after the meeting? Each question has a yes or no answer, with no judgment needed.

The last step is getting tasks out of the minutes and into the places people actually look. The iCalendar standard, RFC 5545, defines a to-do component with a due date, and because not every calendar application recognises to-dos, the same tasks can be written as all-day events instead. Lines in such a file end with a carriage return and line feed and are folded so that none passes 75 octets. A spreadsheet wants CSV, which RFC 4180 describes as quoted fields with doubled inner quotes and the same line endings. Security practice adds one caution: a cell that begins with an equals sign may run as a formula, so a careful export puts an apostrophe in front of it.

Talk time and who gets heard

A transcript holds one more thing that minutes rarely record, which is who did the talking. Caption files carry speaker labels, either as WebVTT voice spans, where a cue opens with a tag naming the voice, or as a plain name and colon at the start of a line. SubRip files, the common .srt format, number each cue and time it as hh:mm:ss,mmm, with a comma before the milliseconds. From those timings, talk time can be measured rather than guessed.

Measuring needs a rule, though, because captions overlap. One speaker’s next cue often starts a moment before the last one ends, and simply adding durations would count that sliver twice. The honest measure is the union: every millisecond in which that person has a cue running, counted once.

When two different people overlap, the opposite choice is the fair one. Both were speaking, so both get the time. The speakers’ totals then add up to a little more than the time anyone was talking, and that difference is worth reporting rather than hiding, because crosstalk is information. A meeting where one person is constantly talked over looks different in those numbers from one where people wait their turn.

None of this identifies a voice. Working out who spoke when from the audio alone is a separate problem, speaker diarisation, which splits a recording into segments by speaker. A measure built on labels is only as good as the labels it reads.

Why measure at all? Because impressions of who dominated a meeting are easy to argue about and hard to check. Turn-taking has been studied for decades in conversation analysis and in gender studies, where early findings that fitted familiar stereotypes have given way to mixed evidence and few overarching patterns. A number will not settle an argument about fairness, but it swaps an impression for a count anyone can check against the recording.

Take the example committee meeting that opens with the tool. It runs from 00:00:04 to 00:04:10, a little over four minutes, with five speakers. The treasurer has four cues of 19.25, 15.5, 4.5 and 13.5 seconds, which add up to 52.75 seconds; none of them overlap each other, so the union is the same. The chair has nine cues totalling 76.0 seconds. There are two short stretches of crosstalk: 1.25 seconds where a member asks a question before the treasurer has finished, and 1 second where two members speak together. That is 2.25 s of overlap, counted for both people each time.

Summed over everyone, the talk comes to 238.75 seconds, while the time covered by at least one cue is 236.5 seconds; the gap is exactly the overlap. The chair’s share is 76 divided by 238.75, or 31.8%. The chair’s longest turn is 26.5 seconds, three cues in a row with short pauses between them, and the pauses do not count, because nobody was speaking.

Keeping every line traceable

In the same example, one action reads “check the hive suits before the open day” and has a due date but no owner. That is exactly the kind of line that drifts. Flagging it at the time, beside the clip of the chair saying it, is cheaper than discovering in late May that everyone thought someone else had it.

The test of good minutes is not elegance. It is whether a doubtful reader can verify any line in under a minute. A decision that carries 00:01:06 and the chair’s exact words can be checked by scrubbing to that point; a paraphrase can only be argued about. The same holds for the files the minutes travel in. A .docx is a zipped package of XML parts, a calendar task can quote the sentence it came from, and both can carry the time in the recording wherever they go.

The limits deserve plain statement. Times are positions in the recording, not clock times. Plain text transcripts often give only start times, so a line is taken to run until the next timestamp, and a pause is charged to whoever spoke before it. And someone who said nothing may still have been in the room: a transcript can show who spoke, never who attended.

Minutes were never meant to be complete. They were meant to be right, and a line tied to its moment is a line anyone can check.