All guides

    Learn / Transcript specifications

    How to Specify Human English Transcription for Speech AI

    By C8C.aiRetrospective series issue:

    This issue date is an editorial label, not a historical publication date or the date a project was performed.

    A useful transcription specification tells listeners exactly what to preserve, how to represent uncertain speech, which optional timestamps to provide, and how spoken numbers become written forms. Agree those rules on a sample and use independent review to resolve ambiguous cases consistently.

    Should the transcript be clean or verbatim?

    Start with the use of the transcript. A speech model may need audible fillers, repetitions, or unfinished words, while a research reading copy may need a cleaner representation. “Verbatim” and “clean” are not complete specifications: teams can interpret both terms differently. Define each category of speech that will be retained, removed, or marked.

    Include rules for false starts, hesitation sounds, self-corrections, and repeated words. Decide whether punctuation represents readable sentences or spoken boundaries. Do not silently rewrite grammar or substitute what a speaker probably intended. A cleaned transcript should still preserve meaning, and the scope should make any permitted edits explicit.

    How should unclear speech and speaker changes be handled?

    Distinguish speech that cannot be understood from sections where no speech is present. Use an agreed marker for inaudible words rather than guessing. If an uncertainty affects interpretation, reviewers need enough surrounding audio to investigate it. Recording quality issues should be logged separately from transcription decisions.

    For conversations, define stable speaker IDs and the treatment of overlap and interruptions. A single-speaker recording can still contain other audible voices; decide whether they are in scope. State how turn boundaries appear in the deliverable. These decisions prevent different listeners from producing incompatible transcripts of the same audio.

    Which timestamp and number-format settings need agreement?

    Optional timestamps are a transcript format choice. Specify whether the handoff needs turn, segment, or word timestamps; the time unit; the reference file; and how uncertain boundaries are represented. Choose the detail the downstream workflow actually uses rather than adding fields simply because they are possible.

    Inverse text normalization converts spoken forms into written forms, such as dates and numbers. Agree conventions for dates, currency, decimals, abbreviations, and quantities. A spoken date may be ambiguous without context, so the rule should say when to preserve the spoken form or flag uncertainty rather than invent a calendar interpretation.

    What can a transcription pilot reveal before expansion?

    C8C.ai worked on single-speaker English transcription pilot material with written timestamp and inverse text normalization rules and an independent review workflow. The general lesson is that a pilot should test the specification as well as the listeners. This is not a description of customer material, measured performance, or when the work took place.

    Compare independent reviews against the same written rules. Classify disagreements: unclear audio, missing guidance, or an inconsistent application of a rule. Adjudication should produce a decision and, when useful, a reusable clarification. Review a fresh sample after updating instructions so that agreement on an earlier example is not mistaken for consistent future application.

    Agree these settings before transcription

    • Reading style: define retained fillers, repetitions, false starts, and permitted cleanup.
    • Uncertainty: specify inaudible and overlap markers and an escalation path.
    • Timing: choose optional turn, segment, or word timestamps and their reference.
    • Normalization: document date, number, currency and abbreviation conventions.
    • Review: arrange independent checking, adjudication, versioned rules and a sample handoff.

    Put the brief into practice

    Bring the intended use, sample requirements, and review questions to a scoping conversation. Agree the deliverable and pilot checks before expanding production.

    C8C.ai / A Focus Forward company