meeting-intelligence

Turns a meeting transcript into decisions, action items and a summary. It reads text, never audio: there is no recorder, no decoder and no speech model in this package, so getting words out of a recording is a separate job done before meeting-intelligence is called. Export captions from your meeting tool (Zoom, Teams and Meet all produce a .vtt file), or use a dedicated speech-to-text package - the speech and audio packages in this family, or any tool that writes .srt/.vtt/plain text - and hand the result to analyse().

Nothing is downloaded, no model is loaded, and every decision and action can be traced back to the published cue phrase that produced it.

Install

pip install meeting-intelligence

Quickstart

import meeting_intelligence as mi

report = mi.analyse("""
Alice: We decided to go with Postgres for the event store.
Bob: I'll write the migration by Friday. Can you review it, Alice?
Alice: Sure. What do we do about the old rows?
""")
print(report.summary())
print(report.action_items[0].owner, "->", report.action_items[0].due)

That is the whole API for most people. Pass plain text, a list of {"speaker", "text", "start"} turns, a path to a .txt/.vtt/.srt/.md file, or a WebVTT or SubRip string - the parser works out which it got and reports it as report.source_format.

What it does

• Decisions come from a published cue list, mi.DECISION_CUES: we decided, the decision is, let's go with, agreed to, the plan is, settled on, signed off on and a dozen more. Each Decision carries the cue that matched and a confidence. A cue inside a negation ("we have not decided", "still deciding") or inside a question is never reported as settled.
• Action items come from mi.ACTION_RULES: I'll take, I will, let me, X will, X owns, can you, Name, can you, assigned to X, please, action item:, follow up, we need to, needs to be, make sure, todo. Each Action is Action(text, owner, due, confidence) and carries the cue that matched. I'll on its own is not a commitment: I'll be honest, I'll say, I'll admit, I'll second that, I'll bet and I'll leave it there are remarks, not tasks, and are not reported as action items.
• Owners are people the report can name. I and I'll become the speaker of the turn. can you becomes the person addressed: a name written into the sentence wins, whether it leads (Alice, can you review it?) or trails (Can you review it, Alice?), and otherwise it is the other person on a two-person call or the next person to speak. A named owner is resolved against the roster - the speakers found in the transcript plus anything passed to speakers= - so a bare first name matches a fuller roster name. An owner that cannot be resolved is left None rather than guessed: a capitalised word that matches nobody is not a person, so Hey, Sorry, Folks and Hopefully Monday never appear as owners. Pass speakers=[...] to name somebody who is mentioned but never speaks.
• Due dates come from mi.DUE_PATTERNS: by Friday, end of week, EOD, by 2026-04-01, within two weeks, ASAP. The leading "by"/"before"/"due" is trimmed, so due reads as a date.
• Questions are matched to whether a later turn answered them. A different speaker replying within six turns counts as the answer when the reply shares a content word with the question, or when a who question is answered by naming somebody ("Assigned to Dana."). A bare affirmation ("yes", "no", "it is", "we can") counts only in the next two turns, where a direct answer actually lands, and a sign-off ("That is everything. Thanks all.") never counts - an unanswered question stays open rather than being paired with an unrelated later turn. report.open_questions is what nobody answered. Tag questions and filler ("Right?", "Okay?") are not counted as questions.
• Topics are segments of the discussion. Each turn becomes a hashed bag-of-words vector, the cosine similarity across every turn boundary is measured with a window on each side, and the discussion is cut at the valleys. The label is drawn from the words that are distinctive to that segment relative to the rest of the meeting - meeting scaffolding, greetings, dates and the cue verbs themselves (decided, agreed, assigned) are kept out so the label names the subject, and an acronym keeps its capitals (PR, not Pr). numpy does the arithmetic.
• Participation is per speaker: turns, words, share of the talking, longest monologue, questions asked and interruptions. Interruptions are counted from timing overlap when the source is timed, and from cut-off markers (--, ...) when it is not; which rule was available is reported as report.interruptions_basis, so the number is never ambiguous.
• A heuristic that finds nothing says so. A transcript too short to split into topics, a transcript with no speaker labels, and a source with no way to detect interruptions each add a line to report.warnings instead of silently returning something degenerate.
• An empty transcript returns an empty report, not an exception.
• Deterministic: the word hash is zlib.crc32, so the same transcript gives the same report in every process and on every machine.
• Unicode throughout - names, text, and CJK sentence enders (。, ！, ？).

Timestamps and formats

Consecutive cues from the same speaker are merged into one real turn, so a caption track that breaks a sentence across four cues still counts as one turn and one monologue.

Using your own LLM (optional)

report = mi.analyse(transcript, llm=my_model)   # callable(prompt) -> str

The llm is used only to reword the summary paragraph and to tidy the phrasing of action items that were already extracted. It never finds new decisions, owners or dates, it is never required, and the package is fully useful without it. If your callable raises, returns nothing, or returns something unusable, the extractive result is kept and the reason is appended to report.warnings. report.llm_used says whether any of its output was used, and Action.source_text keeps the original sentence whenever the wording was changed.

API

mi.analyse(transcript, *, speakers=None, llm=None) -> MeetingReport
mi.parse_transcript(source) -> TurnList        # list[Turn], with .source_format
mi.detect_format(source) -> str                # name the shape without keeping the parse

mi.analyze is the same function under the US spelling.

MeetingReport fields: summary_text, decisions, action_items, questions, topics, participation, duration, plus turns, speakers, source_format, n_turns, n_words, interruptions_basis, llm_used and warnings.

MeetingReport methods and properties: summary() (plain text for a terminal), to_dict() (JSON-safe), to_markdown() (minutes to paste), open_questions, answered_questions, owned_actions, unowned_actions, duration_text, is_empty and actions_for(owner).

Record types, all with to_dict(): Decision(text, speaker, cue, confidence, ...), Action(text, owner, due, confidence, ...), Question(text, speaker, answered, answer_text, ...), Topic(label, keywords, start_turn, end_turn, ...), Participation(speaker, turns, words, share, longest_monologue, interruptions, ...) and Turn(index, speaker, text, start, end).

The pieces are exported individually too, for callers who want one of them on its own: extract_decisions, extract_actions, extract_questions, segment_topics, measure_participation, find_due, unanswered and format_duration - along with the cue lists DECISION_CUES, ACTION_RULES and DUE_PATTERNS.

CLI

meeting-intelligence standup.vtt                          # the text summary
meeting-intelligence notes.txt --speakers "Alice,Bob Chen"
meeting-intelligence notes.txt --json > report.json
meeting-intelligence notes.txt --markdown --output minutes.md
meeting-intelligence notes.txt --actions                  # just the task lines
cat notes.txt | meeting-intelligence -                    # transcript on stdin

meeting-intelligence --help lists every option. Output is UTF-8 even through a pipe, so a transcript full of non-ASCII names prints without a UnicodeEncodeError.

License

MIT