Project
Audio to transcript to events to a checked record, with a source on every fact.
Where it runs: On my home server (private)
The path is: audio, then a transcript, then extracted events, then a checked record. Each step is a place where an error can enter, so each step has to leave a trail back to the one before it.
My test case is an amateur football club that streams almost every match with live commentary. Much of its history (line-ups, substitutions, cards, who scored and how) exists only in what the commentator said on those recordings. There are about 125 streams of about 100 minutes each, roughly 200 hours, and nobody can rewatch that to answer a question. A summary written by an AI model is only useful if you can see where each fact came from and whether anyone has checked it.
I built a pipeline that follows that path and turns the recordings, the club's public match pages and its public posts into one match database with a strict evidence discipline. Every fact stores its source. Anything a language model extracted enters as unverified and becomes verified only when a second source agrees or a person checks it. Totals are computed, unknowns stay empty, names are matched through an alias table, and disagreements go to review lists instead of being resolved silently. A private statistics site and a read-only assistant sit on top, and every extracted event keeps the second in the stream where it was said.
Collect public sources
Match pages from the club and the football associations go into tables with a source address on every row. The other inputs are the club's public live streams and public posts.
Transcribe locally
The Greek commentary is transcribed with an open speech-to-text model on my RTX 5090 workstation. Known errors of the model (music read as captions, repeated loops) are removed by explicit rules, and every removal is logged in the transcript file. The audio is deleted after transcription.
Extract events
For each match, an AI agent reads the full transcript with a fixed written instruction ("only what the speaker says, never guess") and writes goals, cards, substitutions and line-ups as structured data. Each event carries the second in the stream and the short quote it came from.
Load with the evidence rules
A loader stores everything as unverified, checks it against the match pages (same team, minute within 2), marks agreeing facts as verified, and writes unmatched names and conflicts to review lists. Score cards in posts, read by a local vision model, act as a further independent source.
Review and answer
I work through the review lists. The statistics site and the assistant read the database, and both show what is still unverified.
Each goal, card, line-up and score points to a source record: a match page, a post, a stream or its transcript, with the address and the time it was retrieved. A separate table records which source covers what for each match (score, line-up, events), so you can see why a value is there.
The database default is "unverified". A goal extracted from a transcript becomes verified only when the club's match page has the same goal by the same team within two minutes, or a person confirms it. A different scorer does not overwrite anything: it goes to a conflicts list. The site marks unverified items with a question mark and can count verified goals only.
Goals, minutes played, records and head-to-head figures are always calculated from the individual events, never typed in. When the commentator says a running total, it is kept as a "reported" figure to compare with the calculated one. A value nobody said stays empty, not zero. Where minutes played depend on a missing value, the total is labelled a lower bound.
The assistant answers only from tools that read the database, never from memory, and says so when something is not there. It uses the fixed queries first and searches the transcripts only for moments and quotes the database cannot cover. It mentions when an event has only one source and links to the stream about ten seconds before the moment. The same read-only tools also work with a local model on my workstation.
Speech-to-text gets names wrong, and commentators use short forms. Names are stored as heard, then matched through an alias table that handles Greek case endings and similar-sounding forms. A name that does not match, or matches more than one person, is not loaded: it goes to a review list.
Every extracted event keeps the second in the stream, the quote it came from and a confidence value when the commentary was not explicit. Match minutes are calibrated against the kick-off of each half, so the minute the commentator said can be checked against the recording.
A set of Python scripts on my workstation does the collection, transcription, cleaning, extraction and loading into SQLite, with full-text search over the transcript segments. The rules live in three places, so they do not depend on a prompt being followed: the schema (unverified by default, empty means unknown), the loaders (matching, conflicts, review lists) and the views (every total). The statistics site is a small FastAPI app that serves a read-only copy of the database.
One measured detail: transcribing in 15-second windows instead of 30 kept text that the longer window lost (about 15% of it).
I built it by directing AI coding agents under a written rule file that sets the data rules. Research agents collected the public match data with a source address on every row, and one agent per match did the extraction with the same fixed instruction. I set the rules, confirmed identities, resolved conflicts and reviewed the results.
Limits:
Stack
Still under construction
The Tsakmak brand and this website are still being built, so some pages and details are not final yet. Thank you for your patience.