AI Transcripts: Check the Words That Change a Decision

Imagine a client saying, “We cannot approve the artwork until Monday.” The transcript instead records, “We can approve the artwork until Monday.” A summary then turns the exchange into a confident action item. This is a hypothetical example, but it exposes a useful question: which words in your recording carry the decision?

For a creator preparing an interview, or a business owner reviewing a project call, checking an AI transcript means more than correcting spelling. The job is to preserve what the speaker actually said, including conditions, uncertainty and limits. A readable page is a starting point for that work.

An error count cannot weigh a promise

Microsoft’s speech-to-text documentation defines word error rate as the number of substitutions, deletions and insertions divided by the number of words in a reference transcript. It is useful for comparing a machine’s output with a checked text.

The calculation counts word errors without assigning a business consequence to each one. In a hypothetical 100-word passage, deleting a single “not” could produce a word error rate of one per cent while reversing its central instruction. An overall score therefore needs a second question alongside it: did an error change something someone might rely on?

An added sentence is a different kind of problem

In Careless Whisper: Speech-to-Text Hallucination Harms, published on 5 June 2024, Allison Koenecke and colleagues examined Whisper’s behaviour as of 2023. Their sample included 13,140 American English audio clips from AphasiaBank. Roughly one per cent produced hallucinations: phrases or sentences absent from the underlying audio.

The researchers also found a relationship between hallucination and longer non-vocal portions of recordings. Their study included people with aphasia and a control group; it was not a representative test of every workplace conversation. These results establish a failure mode in the studied system and setting. They do not supply an error rate for every transcription product, or for its October 2026 version.

My practical inference is to review unexpected additions as well as obvious mishearings. A surprisingly complete closing sentence deserves checking against the recording just as much as a visibly garbled name.

Start with the words that carry consequences

The following is a proposed review method, rather than a protocol validated by those studies. Before polishing the prose, make a first pass for four kinds of detail:

• Identity: names, organisations, product titles and who said each relevant line.
• Quantity: amounts, dates, units, durations and version numbers.
• Conditions: words such as “unless”, “after”, “only”, “not” and “subject to approval”.
• Commitment: the difference between a suggestion, an intention, an agreement and an unresolved question.

For each consequential passage, listen to enough surrounding audio to understand the exchange. Do not stop at finding a word that sounds plausible. Check whether the next sentence qualifies it, someone interrupts, or the apparent agreement answers a different question.

This first pass prioritises risk; it does not certify the rest of the transcript. Before publishing a quotation, check the complete quoted passage against the audio. Review the whole text when the intended use requires a faithful record throughout.

Keep the recording, transcript and decision distinct

Where an approved recording is available, use it to check what was audible. Keep the working transcript visibly labelled as reviewed or unreviewed, and record consequential corrections with a timestamp. If the software supplies no timestamp, add a manual reference that lets another reviewer find the passage.

Put confirmed actions in a separate decision record. That record can state the owner, deadline and outstanding approval without presenting your interpretation as a quotation. The MaryChuks guide to building a decision log explains how to preserve those choices after a meeting.

A speaker’s later clarification belongs in a dated note. It can settle what the team should do next without silently rewriting what the original recording contained.

Give unclear words somewhere visible to stay

Return to the imagined artwork discussion. Suppose the crucial word remains inaudible after replaying it. Mark the uncertainty near the passage and pause the action that depends on it. Ask the client to confirm whether approval has been given and what Monday refers to. Record that confirmation separately.

A small review note can contain:

• the audio reference and speaker, if known;
• the uncertain word or passage;
• the consequence of choosing the wrong interpretation;
• the person who can clarify it;
• the resolution and whether it changes a downstream action.

If you have only the transcript, be explicit about that limit. You can inspect inconsistencies and request clarification, but you cannot establish what the audio contained. Asking another text model to make the sentence sound right adds an interpretation; it does not recover the recording.

Check whose speech your workflow handles well

A PNAS study published on 23 March 2020 compared five commercial speech-recognition systems using interviews with Black and White American speakers. All five showed substantial racial disparities in word error rates in that dataset. Those historical results are not a ranking of today’s systems.

They support a sensible testing question: whose speech is missing from your trial? Include the voices, vocabulary and recording conditions your work actually involves, with participants’ agreement. Check speaker attribution and consequential errors separately. A small local trial can expose problems; it cannot establish that a system is fair for every group.

Make the correction route easy to use, including for someone who prefers to confirm an action in writing. Count the time this takes when deciding whether automation helps. The separate guide to the workflow around an AI meeting assistant explores that operating cost.

A useful transcript lets a reader follow the conversation and find its limits. When an uncertain word could change a promise, leave a visible question until there is evidence to answer it.

Sources

• Koenecke and colleagues, Careless Whisper: Speech-to-Text Hallucination Harms, ACM FAccT, 5 June 2024. Author-hosted extended summary.
• Koenecke and colleagues, Racial disparities in automated speech recognition, PNAS, first published online 23 March 2020.
• Microsoft Learn, Speech-to-text FAQ, word error rate definition; checked 23 October 2026.