transcript-conditioning

Resolve speakers by content rather than by the diarized labels. Match each voice to a known roster using role tells like company self-references, domain vocabulary, and moderator question patterns, then record confidence per resolution and flag voices that cannot be resolved.

Build a mangled-to-canonical glossary anchored to a pre-event brief listing known companies, people, products, and acronyms. Only correct garbles that map confidently to a brief entity, and exclude unresolvable ones from any public quote rather than guessing.

Condition whenever the transcript comes from a manual paste, phone recording, or unreliable diarization and you plan to quote named people publicly. Skip it when the diarized transcript is already clean and the speaker roster is small and obvious.

No. Quotes tagged MED or drawn from low-signal sections like mic dropouts or crosstalk should be paraphrased and softened, never promoted to verbatim quotes. Only HIGH-confidence lines are safe to quote word for word.

Transcript conditioning is single-event and content-facing, resolving speakers and entities against a known roster to produce a quote bank. Sales call analysis mines corpora of ten or more calls for pains, triggers, objections, and battlecards.