Documentation time
EstimatedReduced
The workflow created structured drafts that reduced manual minutes preparation for recurring meeting formats.
Workflow Automation
Meeting Minutes Generator — Speech-to-Structured-Action Automation
A speech-to-structured-actions system that turns meeting recordings or transcripts into summaries, decisions, and follow-up tasks.
Sanitized case studyMay 30, 2026
The Meeting Minutes Generator converts meeting recordings or transcripts into summaries, decisions, and follow-up tasks that teams can actually use.
Whisper/STTLLMChunkingStructured OutputDOCX/Email Export
Problem
Meetings generate decisions and actions that are easy to lose. Manual notes take time, and action ownership often becomes unclear after the meeting ends.
Users
The users were business teams, project teams, and internal stakeholders who needed clear documentation after meetings without turning every meeting into a manual admin task.
My role
Constraints
- Long transcripts can exceed context windows.
- Audio quality and transcription errors affect downstream summaries.
- Speaker identity may be ambiguous.
- Action ownership and due dates are often implied, not explicit.
- Outputs must be editable, not treated as final truth.
System built
The system accepted audio or transcript input, produced a transcript when needed, split long content into manageable chunks, summarized each chunk, merged summaries, extracted decisions and actions, and exported a structured minutes document.
- 01Audio or transcript intake
- 02Speech-to-text
- 03Chunking
- 04Chunk summaries
- 05Global summary
- 06Decision extraction
- 07Action extraction
- 08Document export
Architecture snapshot
The pipeline separated transcription, summarization, extraction, and export so each step could be evaluated independently.
AI/ML approach
The AI approach used speech-to-text for recordings and chunked LLM summarization for long transcripts. Structured extraction prompts turned discussion into decisions, open questions, risks, and follow-up tasks.
Product decisions
- Output minutes in a predictable template.
- Separate decisions from action items.
- Mark uncertain ownership rather than inventing it.
- Preserve enough context for human correction.
- Export into formats teams already use.
Evaluation & reliability
01 · Grounding
Generated actions and decisions were checked against transcript evidence.
02 · Failure modes
Risks included hallucinated tasks, missed decisions, transcription errors, and unclear speaker attribution.
03 · Human review
Users reviewed and edited minutes before distribution.
04 · Operational risk
The system could create follow-up confusion if it invented owners or deadlines.
05 · Adoption
Adoption depended on producing useful drafts that were faster to edit than writing from scratch.
06 · Production readiness
A mature version needs diarization, confidence flags, prompt versioning, and export audit history.
Impact
Limitations
What I learned
Business automation succeeds when it produces a useful draft in the user's existing format. The AI output must be easy to verify, edit, and share.
What I would improve next
- Add speaker diarization and identity correction.
- Add action confirmation before export.
- Integrate with calendar and task tools.
- Track action extraction precision and missed decisions.
Expandable technical notes
The main reliability pattern is chunk isolation followed by synthesis. Each chunk produces local notes, then a second pass merges them into a global meeting record. A production version would store chunk provenance so every final action can be traced to transcript evidence.