Gold transcripts

Gold transcripts are the reference text your models are measured against. Annotators transcribe the audio, a second pass reviews against a style guide, and any disagreement is resolved by a senior reviewer before the line is frozen. The result is a transcript you can trust to the punctuation mark — a reference that will not silently move underneath your metric between one training run and the next.

The workflow starts with the style guide, and the style guide is yours. The casing policy, the abbreviation policy, the handling of fillers, overlaps, disfluencies, non-speech events, numbers and proper nouns are written down and versioned before the first clip is transcribed. Each clip passes through two independent transcribers; where they agree the line is frozen, where they disagree it is escalated to a senior reviewer who settles the line with a written note that travels with it.

Every line carries timestamps, speaker labels and the event tags your pipeline expects, in the schema you already run. The raw audio, the two passes and the adjudication notes are kept behind each transcript. What you receive is a versioned transcript set: the audio reference, the frozen transcript with timestamps and tags, the style guide version, and the full evidence trail behind every contested line. The point of a gold transcript is not the text; it is the reference that will not move.