Transcript Audit

What this does and why it works

Auto-transcription errors on accented speech are not random. They are phonetic substitutions: the transcriber hears the right sounds and picks the wrong words. In a technical talk this hits the domain jargon hardest, because the transcriber has no model for "HEPES" or "oligo-dT" and falls back on common English ("heapy-space," "oligodity"). Two things make these highly catchable:

1. A domain lexicon. The same terms get mangled the same way across every recording. references/lexicon.md holds the known mishearing → correct-term mappings plus the RNA/bioprocess and Invert terms most likely to be garbled.
2. Context clues. The transcript usually contains its own answer key. A term mangled in one place often appears correctly elsewhere; a defined acronym gets used later; a number is restated. Internal consistency is your strongest tool — lean on it before the lexicon.

This skill produces an audit list, not a cleaned transcript. Nick applies the fixes himself (often in a separate editor). Do not rewrite the transcript unless he explicitly asks — the job is to surface catches he can act on fast.

Method

1. Read the whole transcript first. Do not audit top-to-bottom on first pass. Later passages disambiguate earlier ones (a buffer named "heapy-space" at minute 20 is confirmed by "conjugate bases for HEPES" at minute 40).
2. Cross-check against the lexicon. Read references/lexicon.md and scan for its entries. Treat it as a checklist of usual suspects, not a limit — catch anything that fits the phonetic-substitution pattern even if it's not listed.
3. Run consistency checks. Flag internal contradictions: numbers that don't match a value stated elsewhere, an acronym defined one way and used another, a term spelled two ways. These catch the errors the lexicon can't.
4. Classify each catch by confidence (see Output). The single most important discipline: fix what context proves; flag what you can't verify. Never silently "correct" a proper noun.
5. Leave disfluencies alone. Stutters, false starts, and filler ("the the," "two two two," "um") are left verbatim — Nick keeps the transcript faithful to the recording. Note in one line that they remain; do not itemize them.

Output format

Return a grouped list. Quote the exact heard text in each entry so Nick can find-and-replace directly. Keep entries to one line of "why" — the reasoning is what lets him trust or overrule the catch. Use this structure:

**Technical / domain terms** (high confidence — misheard jargon)
- "heard text" → correction — one-line why (usually the context clue)

**Meaning errors** (wrong word or dropped word that changes the meaning)
- "heard text" → correction — what it does to the meaning

**Verify — can't confirm from the text** (names, products, instruments, figures)
- "heard text" → best guess (if any) — why it's uncertain + what to check against

**Punctuation / run-ons** (only if egregious or if Nick asks)
- brief note

**Disfluencies**
- One line: left verbatim per your default; note they remain if a clean read is wanted.

If a section has no catches, drop it. Lead with the substantive tiers (Technical, Meaning) — those are what Nick acts on first.

The two rules that matter most

Context is the answer key. When a term appears both mangled and correct in the same transcript, the correct instance is ground truth: fix the mangled ones to match and note the inconsistency. When a number or claim contradicts one stated elsewhere, flag the contradiction and say which value the surrounding context supports. These internal-evidence catches are more reliable than any lexicon and are the ones a human skim misses.

Flag, don't guess, on anything you can't verify. Speaker and company names, product names, instrument makes/models, citations, and specific figures cannot be recovered from audio-driven text alone. Offer a best guess when you have one, but mark it clearly and tell Nick what to check it against (his slides, notes, the attendee list). Silently "fixing" a name is worse than flagging it — it launders an error into something that looks confident.

Worked examples

Example 1 — context-clue fix (Technical): Input: "commonly used buffers in IVT are heapy-space or tree-space buffers" Output: "heapy-space or tree-space" → HEPES or Tris — later the speaker says "conjugate bases for HEPES" and "we can do it for Tris," which confirms both.

Example 2 — consistency catch (Meaning): Input: "40 species for the full model and 50 species for the simplified model" Output: "50 species for the simplified model" → 15 — everywhere else the simplified (Henderson-Hasselbalch) model is 15 species and the kinetic model is 40; "50" contradicts the rest of the talk. Worth confirming against the slide.

Example 3 — flag, don't guess (Verify): Input: "we just used a tiny pH probe, a five-rotic pH probe" Output: "five-rotic pH probe" → likely "fiber-optic pH probe" — fits "tiny" and fiber-optic pH microsensors exist, but it's a specific instrument that can't be confirmed from the text. Check the slide/notes before trusting it.

Example 4 — dropped word (Meaning): Input: "a lot of automation which would / the number of operators needed" Output: insert "reduce" → "automation which would reduce the number of operators" — a verb dropped out; the sentence is incomplete without it.

Growing the lexicon

The lexicon is the compounding asset here — each recording surfaces a few new mishearings worth saving. When Nick confirms a new mapping (or gives you Invert product/people/competitor names to watch for), offer to add it to references/lexicon.md. If this skill is installed and read-only, note that he can ask you to regenerate an updated .skill package with the additions.