Before an AI agent turns an interview into an article, check the transcript against the recording. A fluent sentence can contain the wrong amount, a missing “not” or a claim assigned to the wrong speaker. Those errors can survive every later editing pass because the text itself looks ordinary.
This reference offers sixteen checks for transcripts used in research, publishing and content production. It does not rank speech models. The question is whether a particular passage supports the use you plan to make of it, including what remains unclear or absent from the recording.
- 01Replay consequential passages.Check quotations, names, numbers and conditions before they become published claims.
- 02Separate speakers from identities.A consistent speaker label does not establish a person’s name.
- 03Preserve uncertainty.Mark unclear or missing material instead of filling it from a plausible story.
01 — Check four layers of evidenceCheck four layers of evidence
Words cover what was said; speakers cover who said it; timing covers where it appears; context covers what text alone may omit. A check can be not applicable: audio-only material has no visual track to inspect. An unavailable recording is unverified, not a clean result.
| Review area | Evidence to inspect | Editorial consequence |
|---|---|---|
| Words: Names and specialist terms | Replay the phrase with its surrounding sentence | Correct from audible evidence or a verified glossary; preserve uncertainty otherwise. |
| Words: Numbers and units | Check each consequential quantity against the audio | Do not silently turn an unclear amount into a precise price. |
| Words: Negation | Replay statements containing no, not or exceptions | A missing negation can reverse the speaker’s meaning. |
| Words: Spoken versus translated | Identify the language and transcription task | Label a translation as a translation; keep source wording when needed. |
| Speakers: Turn boundaries | Compare speaker changes with the recording | A sentence crossing a turn boundary may belong to two people. |
| Speakers: Named identity | Check an explicit introduction or trusted recording context | Use an anonymous label when identity is not established. |
| Speakers: Overlapping voices | Replay simultaneous speech | Preserve overlap or an unclear span rather than inventing a clean dialogue. |
| Speakers: Quoted speech | Check whether the speaker is quoting someone else | Do not attribute a quoted position to the person reading it aloud. |
| Timing: Missing opening or ending | Compare transcript coverage with the full recording | Disclose clipping rather than implying the transcript covers the whole source. |
| Timing: Timestamp alignment | Jump to selected anchors in the source | Make navigation land near the represented speech; record any offset. |
| Timing: Repeated segment | Compare repetition to the corresponding audio | Remove decoding repetition only when the recording does not contain it. |
| Timing: Silence rendered as speech | Listen to the interval supporting suspicious text | Delete unsupported words and retain an uncertainty note where appropriate. |
| Context: Meaningful non-speech sound | Check laughter, alarms or other relevant audio cues | Include cues that change interpretation without inventing emotions. |
| Context: Visual-only information | Inspect the video if a descriptive transcript is required | Describe relevant visible information separately from spoken words. |
| Context: Editorial cleanup | Compare the edited transcript to the source version | Label summaries and clarifications; avoid passing paraphrase off as verbatim speech. |
| Context: Unclear passage | Retain a time anchor and the unresolved span | Mark uncertainty instead of completing the thought from context. |
02 — What the primary guidance supportsWhat the primary guidance supports
W3C’s transcript guidance discusses speaker identification and the visual information a descriptive transcript may need. It also distinguishes clearly marked additions from the original audio and treats timestamps as an optional aid. These are useful editorial boundaries, not a universal transcript layout.
The Whisper model card warns that output can include text absent from the audio and that performance varies across languages and accents. This is a limitation documented for that model family. It does not provide a present-day error rate for every transcription system.
Our checklist turns those distinctions into review questions. We did not transcribe a corpus, measure a model or test demographic differences. The table’s counts describe the editorial reference only.
03 — Make an unclear passage recoverableMake an unclear passage recoverable
For an illustrative interview, suppose an amount sounds like either fifteen or fifty. The surrounding conversation might make one more plausible, but plausibility is not audio evidence. Preserve a time anchor and mark the amount unclear. If the amount matters to the article, obtain clarification or omit the unsupported precision.
Keep the original transcript, the reviewed version and the recording identifier connected. That lets an editor return to the exact passage after a correction. A timestamp should locate evidence, not merely make the document look technical.
Speaker labels need the same discipline. “Speaker 2” can remain useful when a name is unknown. Do not attach a known executive’s name simply because the statement sounds like something that person might say. The research citation reference applies the same principle to written attribution.
04 — Keep quotation, cleanup and summary distinctKeep quotation, cleanup and summary distinct
A verbatim passage aims to preserve the spoken wording. A cleaned transcript may remove fillers under an agreed editorial policy. A summary rewrites the material around its meaning. An agent should state which output it is producing so a later editor does not put quotation marks around a paraphrase.
For a video with slides, decide whether the deliverable needs the visible material as well as the speech. If a speaker says “this figure,” the audio may not identify the number. Describe the relevant visual information separately and retain its source location.
Translated transcripts add another transformation. Preserve the original-language passage when a quotation or commercial promise depends on precise wording. Our translation quality guide covers checking meaning-bearing details through that change.
05 — Review the claims the transcript will supportReview the claims the transcript will support
Review depth should follow the intended use. For search within a recording, rough text with visible uncertainty may be useful. For a published quotation, replay the complete passage and enough surrounding material to establish its context. Do not claim that a small spot check validates an entire recording.
A handoff should say which intervals were checked and which were not. If the recording cannot be recovered, the transcript may still support navigation or a tentative lead, but the missing evidence must remain visible to the next editor.
Keep the transcript readable as well as accurate. The accessibility verification article explains why visual polish alone is an incomplete review of a content artifact.
Original editorial classification informed by inspected primary sources.
- What was collected
- 16 transcript checks across 4 evidence layers. One row per check or response case; multiple rows can apply to the same artifact or task.
- As-of date
- Sources inspected and classification assembled September 7, 2026. The assigned publication date is September 5; this is not a claim of collection on that date.
- Sources and evidence
- The linked primary documents provide conceptual anchors. Row selection, grouping and recommended actions are Digital Applied editorial analysis, not a checklist issued by those sources.
- Counting method
- Count each table body row once and group by the label before its colon. The chart renders those counts with a common linear scale; units are checklist entries.
- Exclusions and gaps
- No vendor census, model ranking, live experiment or prevalence estimate. UNVERIFIED means the required evidence was not inspected. Not applicable means the property is outside the agreed task.
- Limitations and updates
- This is a bounded reference, not an exhaustive standard or certification. Review the same URL when source guidance or use cases change; a passed checklist does not establish every aspect of correctness.
06 — DecisionWhat to do next
Let the recording settle the consequential words.
Choose the passages your next output depends on, replay them and preserve unresolved spans. The transcript becomes useful evidence when its boundaries are visible.
For implementation support, explore our AI transformation services.