There's a lot of audio sitting unleveraged in organisations, even where the first step
from audio to text has already been taken. The premise of this project is to turn hours
of recorded speech into a graph you can actually ask questions of, with every answer
traceable to who said it, and when. I built a prototype for it in 2024, shelved it when
life got in the way, and rebuilt it in 2026 to run entirely on a single laptop. If you're
sitting on recorded calls, interviews, panels or podcasts, you probably have some kind of
full-text search over the transcripts - but that is usually where it ends. This is about
going further.
The corpus throughout is Mikroökonomen, a German
podcast on economic and political developments. The 12-episode proof-of-concept (494
chunks) used for most of the worked numbers is one slice of a ~460-episode archive, and
the 352-episode core format is what the case study processes end-to-end.
The stack is entirely local - mlx-whisper → pyannote → GLiNER2 → a small local
instruct model → Neo4j - and every stage is cached, so iterating is cheap and a rebuild
only costs you where the pipeline actually changed. The chapters that follow take each of
these stages in turn.
One 23-second clip of Mikroökonomen 237, carried through every stage: audio, transcript, who said it, the graph, a cited answer.
How this was made. The pipeline and the articles you're about
to read were both created with AI assistance. Responsibility for both, and
especially for the statements made in the articles, lies with me. I acknowledge
and respect that many people, for ethical or other reasons, will not want to
consume content that was created in part or wholly using LLMs. This is your notice.
My intent is to inform, not deceive.