How Unspoken Words Enter AI Meeting Transcripts

A meeting assistant writes, “We approve the renewal.” The sentence appears in the transcript, then in the summary, then beside an assigned task. Nobody clicks the recording. If those words were generated during a pause, each downstream document has made the same unsupported statement look more established.

This is a constructed example, not a ZharfAI customer incident. It frames a practical decision for teams using recorded meetings, interviews and support calls: which transcript passages may become evidence of what happened, and which require replay or an explicit unresolved label? The answer starts with the audio, not the fluency of the text.

A transcript makes several separate promises

“These words were spoken” is different from “this person spoke them,” “they were spoken at this time” and “they authorize this action.” A transcription system can be right about one and wrong about another. Even a perfectly transcribed quotation does not determine whether a participant had authority to approve a purchase.

Separate literal transcription from edited minutes. The first attempts to represent audible speech. The second selects, paraphrases and interprets it. Translation adds another transformation. Keep those products identifiable; a polished English summary of a Persian discussion is not an English verbatim record.

ZharfAI's proposed operating rule is to retain a traceable audio span for each consequential statement and keep uncertainty attached until it is resolved. This is an engineering recommendation, not a certification scheme. It does not make a general-purpose recognizer suitable for medical, legal or other high-stakes decisions.

Do not solve uncertainty by asking another language model whether the sentence sounds plausible. Plausibility is precisely what makes an invented sentence easy to reuse. A reviewer needs access to the relevant sound and surrounding context, with permission appropriate to the recording.

What the evidence establishes—and what it does not

OpenAI's Whisper model card explicitly warns that output can include text absent from the audio, repetitive passages and uneven performance across languages. It recommends against high-risk decision-making uses. The card describes limitations of that model family; it does not supply today's error rate for your meeting product.

The 2024 paper Careless Whisper reported hallucinated sequences in 1.4% of 13,140 English audio segments in its April–May 2023 experiments. Longer non-vocal portions were associated with greater occurrence. Its AphasiaBank sample, historical service configuration and detection method matter: the percentage is not a forecast for Persian meetings or every current recognizer.

A separate ICASSP 2025 study tested Whisper large-v3 on non-speech sounds and found recurring invented outputs. Its authors investigated filtering characteristic phrases. That supports testing non-speech inputs, not treating a phrase blacklist as proof of accuracy. A phrase can also occur in real conversation.

The operational inference is narrower than “AI transcription cannot be trusted.” Evaluate absent-speech behavior alongside ordinary recognition, and keep a path back to the recording. No percentage from these experiments establishes a universal acceptance threshold. ZharfAI has not run a new model comparison for this article.

Keep the original clock when preparing audio

Store the authorized original recording separately from working derivatives. Associate it with a stable recording identifier, integrity hash, duration, channel description and access policy. A hash helps establish which bytes were reviewed; it does not establish that the recording itself is authentic or complete.

Resampling, denoising and silence removal should produce identifiable derivatives, not silently replace the source. Record the transformation, version and mapping from each retained segment back to original time. If the recorder has gaps or dropped packets, mark those as capture failures rather than ordinary conversational silence.

For example, suppose a 60-minute file loses a cumulative 900 seconds before original time 00:35:00. That point appears at 00:20:00 in the compacted derivative. A reviewer sent to minute twenty of the original will hear the wrong passage. Keep an interval map; a single global offset is insufficient when deletions occur at different positions.

Preserve channel separation when the capture system provides meaningful channels. Verify their actual routing with a known recording rather than assuming “left means customer.” Never expand retention indefinitely merely to make an audit convenient. Where source audio cannot be retained, state the resulting limit on later verification instead of describing the transcript as independently replayable evidence.

Silence detection is a filter, not a truth test

Voice-activity detection estimates where speech is present. Segmentation decides which portions travel together into recognition. Neither establishes the words a person said. Removing clearly non-speech intervals may reduce unnecessary recognition work, but aggressive trimming can discard quiet words, sentence beginnings and short negations.

Microsoft's speech-recognition documentation describes the segmentation tradeoff: a longer silence timeout permits longer pauses but delays results; a shorter one can split a single phrase. These are product-specific controls, not evidence that one timeout eliminates hallucinations across recognizers.

Test complete utterances with pauses, not just isolated clean sentences. Include low-volume speech, room noise, hold music, clipped starts, overlapping talk and recordings with no speech. Choose padding and segmentation settings from those tests. Retain the pre-trim source so a reviewer can discover what a filter removed.

Treat a disagreement between the speech detector and recognizer as a review signal: text generated over an interval labeled non-speech deserves attention, but the detector may be wrong. Do not automatically delete everything it rejects. Otherwise a system meant to suppress invented words may erase a quiet speaker's real contribution.

For Persian, keep spelling normalization separate from acoustic judgment. Whether a half-space or character variant is normalized says nothing about whether a word was audible. The broader Persian speech-interface guide covers language, names, numbers and repair; this workflow adds source-backed acceptance of individual passages.

A speaker label is not a person's identity

Amazon Transcribe's channel-identification documentation describes separate transcription of two audio channels, including channel labels and timestamps. Its speaker-diarization documentation instead describes speaker-group labels such as spk_0. These outputs answer different questions: which recorded channel, or which estimated speaker group?

Neither label alone proves a named person's identity. A shared conference-room microphone contains several people; a single participant can change devices. Treat the association between a participant and a label as separate evidence, with its own review status. Do not ask the recognizer to infer demographic or personal attributes to fill that gap.

When speakers overlap, allow “speaker uncertain” even if some words are clear. Avoid assigning a consequential sentence to the nearest confidently labeled turn just to keep minutes tidy. Keep channel and speaker fields distinct, and make the source span replayable without exposing the whole recording to every downstream reader.

Timestamps also need inspection. A timestamp is a useful pointer, not proof that every displayed word aligns precisely to that interval. Provide a small amount of surrounding context and let reviewers expand it when sentence boundaries are unclear.

A worked meeting fragment shows the release decision

Consider this invented renewal discussion. The times refer to the original recording, not a silence-compacted copy. The proposed outputs are test fixtures, not observed behavior from a particular service.

The first problem is a recognition error with a major consequence. The second is invented speech. The third combines attribution failure with a summary inference. Calling all three “low confidence” hides the different repairs they require.

After review, the minutes might record: renewal was not approved in the first passage; options were requested for Friday; another offer to send options was audible but not confidently attributed. They should not turn the silence into a vote or the offer into a purchase instruction.

If the business wants approval, obtain it through the established approval process. Our guide to human approval design explains how to bind a person's decision to a specific proposed action. A transcript can support that process; it should not silently replace it.

Choose a disposition instead of forcing a clean sentence

Use a small set of states that downstream summarizers must preserve. The following is a proposed working policy, not a claim that any vendor implements it automatically.

Retain the raw model result for controlled troubleshooting when policy permits, but do not make it the public transcript after correction. Each accepted passage should carry the source recording, original interval, channel if available, model/configuration version and review record. A summarizer should receive acceptance states along with text, rather than one flattened paragraph that makes uncertainty disappear.

For exact quotes or consequential commitments, direct replay should be part of the workflow. Lower-risk searchable drafts may use sampled review, provided their draft status remains obvious and users can inspect the source. “Accepted” describes the scope of the review performed, not infallibility.

Measure omissions as well as invented words

Evaluate the whole capture-to-minutes pipeline on recordings you are authorized to use. Include actual deployment languages and acoustic conditions. Keep speakers and recordings used to tune thresholds separate from the final evaluation set. Where speaker-group metadata is necessary, collect it appropriately; do not infer sensitive characteristics from voice.

Ordinary word error rate counts substitutions, deletions and insertions relative to reference words. It is useful, but an aggregate can hide the operational significance of losing “not.” For a no-speech clip, the usual per-clip denominator is zero. Report invented words per non-speech minute and the share of non-speech clips with lexical output separately, with clear annotation rules.

Add critical-statement errors, wrong-speaker attribution, missing negations, timestamp navigation failures and reviewer correction time. Measure deletion of real quiet speech alongside reduction in spurious text. Otherwise a filter that deletes everything can look excellent on the metric it was designed to improve.

A sensible acceptance suite includes both “no words should be produced here” and “this quiet phrase must survive.” Label uncertain references honestly rather than forcing annotators to invent ground truth. Report denominators, test conditions and unresolved cases. Set release limits according to use and consequence; this article offers no universal safe error percentage.

Corrections must reach the minutes and assigned work

Keep a dependency from accepted passages to summaries, extracted commitments and tasks. When a reviewer removes an invented approval, invalidate every artifact relying on it and ask the responsible person to review any action already taken. Editing the transcript alone does not retract an email or undo a renewal.

Distinguish correction from deletion. A correction changes what the organization says was heard; a deletion request or retention expiry changes which artifacts may remain. Our guide to AI deletion dependencies explains why audio, derivatives, transcripts, indexes and summaries need a coordinated lifecycle. Retain only what the applicable policy permits.

Revisit the evaluation suite after changes to recording devices, codecs, recognizers, language settings, segmentation or summary prompts. Sample accepted passages too, not only flagged ones. Watch whether reviewers repeatedly discover unsupported statements that the current rules never surface.

The useful stopping point is not a beautifully complete transcript. It is an honest record that distinguishes audible words, unresolved passages and later human decisions. A blank or uncertain span is sometimes the most accurate thing the system can publish.

Source notes — reviewed September 20, 2026

• OpenAI, Whisper model card. Rolling documentation; original models released September 2022, listed turbo release September 2024. Limitations, not a current product-wide error estimate.
• Koenecke and colleagues, Careless Whisper. FAccT, June 3–6, 2024; the cited experiments ran in April–May 2023 on English segments.
• Barański and colleagues, non-speech hallucinations. Submitted January 20, 2025; ICASSP 2025. Controlled Whisper large-v3 experiments; full text reviewed.
• Microsoft, recognizing speech and handling silence. Updated June 5, 2026. Product-specific segmentation guidance.
• AWS, channel identification. Undated rolling documentation, accessed on the review date.
• AWS, speaker diarization. Undated rolling documentation, accessed on the review date. The workflow, decision tables and invented meeting fragment are ZharfAI analysis, not vendor certification or measured customer results.