From the Lab

Transcript Pipeline

meetings → searchable notes, privately

whisperpyannotelocal-llm
$ cat problem.md

The Problem

Meetings produce hours of audio and almost no usable record. Generic transcription hands you a wall of text with no speakers and no structure — technically a transcript, practically useless a week later when you're trying to remember who agreed to what.

$ cat solution.md

The Approach

A pipeline that runs on my own hardware and turns audio into something you'd actually reread:

  • Capture the audio
  • Transcribe it
  • Split the speakers apart
  • Match voices to names so it's not just "Speaker 1"
  • Clean it into structured markdown — searchable, skimmable notes

It runs locally and on a schedule, so the record builds itself without anything leaving my machines.

$ cat impact.md

Why It Matters

This turns ephemeral meetings into a durable, searchable knowledge base — privately, which matters when the conversations are sensitive.

The pattern underneath is reusable for any stream of content: capture → enrich → structure → store. Point it at audio, support calls, or any messy input and you get the same payoff — something findable instead of something lost.

THE RECEIPTS

Capture → enrich → structure → store

audio  →  transcribe  →  split speakers  →  match voices to names
                                                  ↓
                                    structured markdown notes (searchable)

Each step adds a layer the raw transcript doesn't have. The result is notes you can search months later — not a wall of unattributed text.

← Back to the Lab