What transcripts add
to topic detection.
We tested whether the full discussion gives TopicalBoost more useful information than a recording's title and description. The transcripts add relevant detail. They also expose topic matches that need review.
Walk through the examplesPrivate storage and analysis
All five corrected transcripts are saved in the backend. All five analysis runs completed. Our checks found no transcript text on the public fedsoc.org pages. Copies are available below for this review.
Topic accuracy and public markup
Some names map to unrelated entities. The updated podcast results are not appearing in the linked event pages' topical markup.
Three runs for each recording
These are detected-topic counts.
More topics does not mean higher accuracy.
01 · Description onlyA fresh analysis without a transcript.
02 · Initial transcriptAssemblyAI with subject and name context.
03 · Corrected transcriptLLM review, verified names and selective audio reruns.
| Recording | Description only | Initial transcript | Corrected transcript |
|---|
Choose a recording
Each example shows useful additions,
name corrections and unresolved matches.
What this test establishes, and what it does not
We compared a fresh page-only run, an initial-transcript run and a corrected-transcript run on each of five podcast records. Three existing transcripts were reused; two recordings were newly transcribed. Eight short audio reruns supported 13 narrow spelling corrections.
The full corrected texts matched the saved backend fields after reload. The fedsoc.org public-page check searched five distributed transcript phrases per recording in anonymous HTML. This does not establish privacy across every feed or API, or prove that the complete transcript reached TopicalBoost.
The topic lists below are the observed outputs, including errors. "Added" means absent from the page-only run. "Removed" means absent from the corrected run. Renamed labels can therefore appear as an addition and removal even when they refer to a similar concept.
We have not completed independent listening-based accuracy scoring. Each variant ran once, so differences between the initial and corrected runs may include model variability. This test does not establish an SEO lift or caption-ready transcription accuracy.
This is a static review of the October 8 test. Each example includes the full corrected transcript used in the analysis, with a link to the recording page on fedsoc.org. Publishing these copies on this review page does not change the Federalist Society website.