๐๏ธ how the archive learns to hear
Every recorded board meeting on this site comes with a transcript where each line carries a name. Getting the words right turned out to be the easy part. Getting the names right is the work, and this page is the honest account of how we do it, what went wrong, and how we now grade ourselves.
Behind this text: one glowing cloud per voice in the room. Every so often two of them drift together and become one blob. Pulling them back apart, correctly, is the whole job.
Recordings are transcribed by AssemblyAI's Universal‑3.5 Pro, the best model its pre-recorded service offers (we re-check: on November 13, 2026 it still accepted only that model and its predecessor). A speech model on its own has never heard of Boulan Park or Matt Haupt, so before it listens to a meeting it is handed up to 1,000 key terms: a curated roster of trustees, administrators, schools and acronyms, plus every proper noun mined from that meeting's own board packet: the vendors on the bid tab, the presenters, the students named in a resolution.
The November 10, 2026 workshop, start to finish:
Cost of the transcription itself: about ninety cents. The expensive part is everything after.
The speech model does not know who anyone is. It hears voices, groups them into clusters, and labels them A, B, C. Turning "Speaker F" into a person happens afterward, and every step of that can fail quietly. Three failure modes cover nearly everything we have found:
A transcript that says "Unidentified speaker" is honest. A transcript that puts a trustee's name on someone else's words is worse than no name at all, and it reads just as confidently.
No single signal is trustworthy, so a name has to survive several that fail in different ways:
At the November 10 workshop each signal caught something another missed. Content and the automatic namer agreed one cluster was a trustee, Matt; eleven camera close-ups showed it was the trustee sitting next to him. Then the camera and that meeting's own voices both missed that another cluster held two trustees. The voice library caught it: one woman's lines matched her voice from earlier meetings at 0.84 to 0.98, and the other trustee's at 0.48.
Re-running the same check does not make it more right. Adding a check that fails differently does.
For the November 10 workshop that meant scoring every one of the 875 lines against the voice library: 478 confirmed, 17 moved to the person whose voice they clearly were, and the rest checked by context. In the end 2.4% of the words (mostly "okay," "thank you" and "sure") are unidentified. That number is not a failure. It is the part we refuse to guess.
A method you cannot measure is a method you cannot trust. So we hid the answers: for each of four meetings settled line by line by hand, reference voices were built from the other three, every line was scored blind, and the result was compared with the human answer key (2,449 lines in all).
Lower the bar to 0.50 and it names more lines (72%) but starts mistaking guests, people with no reference voice like a visiting presenter, for board members almost six percent of the time. That trade-off is why the bar sits where it does, and why it is adjusted per person: each regular's own lines score between 0.69 and 0.82 against their reference, so a "match" far below someone's usual score is treated as a stranger. (One reference turned out to be unusable, its own owner scoring 0.37 against it, and it is excluded entirely.)
The one false alarm was itself instructive: a voice that "matched" a trustee who, the video showed, was a muted Teams tile at the time. A reference built from six lines is not a reference.
Then we pointed the same method at the 32 meetings transcribed between January 2025 and August 2026, most of them named by the automatic step and spot-audited at the time. It turned out the automatic namer had often put the wrong name on a whole cluster: a board president and vice president swapped at an organizational meeting; recognitions presenters labeled as board members; outside trainers labeled as trustees; a cabinet member's voice under a trustee's name.
Every proposed change is held to the same rule: a cluster is renamed only when a clear majority of its lines match one person's voice above that person's and that meeting's bar. Ten renames fell short and were held back, and we checked the held ones on video and by content. Three would have turned a visiting presenter into a trustee, and the bar was right to stop them. Three others were real errors the bar was simply too cautious to fix (among them the "trustee" whose cluster said "athletics" 68 times), and those were confirmed by hand, with the evidence written down. One could not be settled either way and stays unidentified.
A day later the check reached the one meeting in that span that had been settled by hand, July 22, and it was wrong too. A three-minute update on summer construction at Athens, Smith and Boulan Park had been credited to the maintenance director. The superintendent had introduced it as Mrs. DiPilato's update, her voice matched at 0.92, and the camera showed her presenting. The automatic namer had said DiPilato; the hand check in July overruled it.
A hand check is one more signal, not the last word.
Published November 13, 2026, with the July 22 correction on November 14, along with 12 more lines the corrected library could now place and a public commenter's name that the November 13 republish had dropped from August 18. The corrected transcripts are live on every meeting page and in the downloadable files; the YouTube caption tracks follow, the last three queued for November 15 (YouTube rations caption uploads). Every meeting's previous version is kept, and every change is recorded with its evidence in the repository.
All of it is in the open: the
transcription/
folder holds the pipeline, the back-test (backtest_voice.py), the cluster check and the fix
proposals, and every meeting's speaker evidence, with the method written up in
docs/TRANSCRIPTION.md.
โฉ Back to the archive ยท ๐ณ๏ธ The Singularity, for how the rest of it works
โ๏ธ github.com/akarpo/tsd-boarddocs
No voices were harmed in the separation of these clusters.