Concepts
How cleanup works
How a recording becomes a clean transcript, and why the AI never changes what was said.
The steps
-
Cut at pauses
The audio is cut into chunks of up to 28 seconds at the quietest moment, so cuts fall in pauses. Each chunk is transcribed with the course glossary as Whisper's prompt.
-
Flag hallucinations
Segments that are likely made up, such as text in silence, loops and echoes of the glossary, are flagged. They're left out of the transcript but kept, marked with ⚠, in the raw output.
-
Build paragraphs
Segments are merged into paragraphs at pauses. Corrections from the glossary, like
hetero skedasticity → heteroscedasticity, are applied word for word. -
Optional: the AI step
Claude Haiku sees the numbered paragraphs and the slide text, if any. Paragraphs where Whisper was unsure are marked. It returns only a list: section titles with optional slide pages, term fixes, topics and new glossary terms.
-
Optional: the summary
A free-form summary, clearly labelled as written by AI, with timestamps pointing back into the transcript.
How term fixes are checked
A suggested fix is applied only if the original words occur in that paragraph, the two spellings are similar, and the replacement is either:
- a term from the course glossary, or
- words that appear exactly on the slides, while Whisper's words don't.
Where Whisper was confident, the spellings must be much closer. Otherwise ordinary words from the slides could flip the meaning, like converge into diverge.
Finally, a word-by-word comparison of the whole transcript must show only those fixes. If anything else changed, all of the AI's edits are discarded.
Example
| Suggested fix | Result |
|---|---|
| hetero skedasticity → heteroscedasticity | Applied: in the glossary. |
| Gauss Mark of → Gauss–Markov | Applied: on slide 14. |
| we kind of assume → we assume | Rejected: changes what was said. |
Applied fixes, with their basis, and rejected ones, with the reason, are listed in the lecture's segments.json.
What the AI writes
Section titles and slide references are the only AI-written text in a transcript, and its properties say so. The summary is a separate file. The words of the lecture always come from Whisper.
To redo only the AI step, for example after adding slides or glossary terms, run stenny refresh. It reuses the stored segments and takes seconds. --no-summary keeps the existing summary, and --no-llm rebuilds the transcript without AI. The raw output is never touched.