How to Clean Up a Transcript: Fix AI Transcription Errors

To clean up a transcript, save the raw version, decide how literal the result must be, then correct it in passes while listening to the audio: first the obvious failures, then the words that matter most (names, numbers, dates and negatives), then speakers and paragraphs, and finally readability. Mark what you cannot hear instead of guessing.

An AI transcript can read fluently even when it is wrong. Cleanup should make the record easier to use without hiding what remains uncertain.

Before you edit

Keep the raw output. Save the untouched transcript as a separate file for the period your retention rules allow. Edit a copy so you can compare changes. Restrict access to both copies if they contain private information.

Decide on the standard. Pick one and stick to it:

Follow the project's agreed standard. A research protocol or formal record may require details that a readable web transcript omits. The W3C's transcription guidance distinguishes detailed verbatim records from web captions that can omit non-substantive speech. Do not apply one editing rule to every use.

Know who will read it. A transcript for your own notes needs less polish than one that will be quoted, published or kept as a record.

Pass 1: catch the obvious failures

Skim the whole transcript quickly before fixing individual words. Look for:

• Garbled text. Check the selected language and listen to the source. If the language is wrong, retry a short excerpt with the correct choice. If it is right, compare the model settings and look for quiet speech, noise or overlapping voices before rerunning the whole file.
• Repeated lines. Compare repetitions with the audio before deleting them; the speaker may genuinely have repeated a phrase.
• Missing passages. Compare the final timestamp with the recording's length, then listen through the recording for gaps within it. Meetings Brief works through a long file about five minutes at a time, so include the section joins in your review. A final timestamp alone cannot show that the middle is complete.
• Added text. Check sentences against the audio, including quiet passages. The Careless Whisper research documents invented phrases in the Whisper system it studied. That establishes a risk to review, not a measured error rate for your recording or this browser tool.

Pass 2: fix the words that matter most

Play the recording and correct the transcript as you go. Spend most of your attention on the words that cause real problems when they are wrong:

• Names of people, companies, places and products.
• Numbers, including money, quantities, percentages and dates. "Fifteen" and "fifty" are easy to confuse.
• Negatives. "Can" and "can't", "did" and "didn't" can reverse the meaning of a decision.
• Commitments. Who agreed to do what, and by when.
• Technical terms and abbreviations used in your field. For lectures and talks, check them against the slides; the lecture transcription guide explains how.

Slow the playback for difficult passages, and use headphones.

Pass 3: speakers and structure

• Add speaker labels at each change of speaker. Meetings Brief's file transcriber does not identify speakers. If attribution is uncertain, mark it instead of assigning the words to a person.
• Start a new paragraph at each speaker change and at each new topic within a long answer.
• Add timestamps before passages you may need to find again, such as quotes, decisions and disputed points.

Pass 4: readability

Apply the standard you chose. For clean verbatim:

• Remove fillers and false starts that don't change the meaning.
• Add punctuation so sentences read naturally.
• Keep the speaker's own words and word order. Don't fix grammar in a way that changes tone or meaning.
• Keep words that carry meaning, even if they look like fillers. "I mean, it's fine" and "it's fine" can say different things.

Mark what you can't hear

Never fill a gap with what you think was said. Use consistent markers:

• [inaudible 00:12:41]: you cannot make out the words.
• [unclear: "Delaney"? 00:13:05]: you have a likely word but aren't sure.
• [crosstalk 00:15:20]: people speak over each other and you can't separate them.

Search for these markers at the end and give them a final listen.

Keep a short correction log when meaning changes

Punctuation edits need not become a separate project. For a decision, amount or disputed statement, record why you changed it:

These examples are fictional. A speaker's later clarification can be recorded as a dated follow-up; do not silently present it as words heard in the original recording.

Copy this transcript style sheet

A short style sheet keeps several editors, or the same editor across several sessions, consistent.

---

Transcript style sheet: [project]

Standard: [verbatim / clean verbatim / edited]

Speaker labels: [e.g. full first name in capitals, or INT / P1]

Timestamps: [format, e.g. 00:12:41; where they go]

Spellings: [names, products, places and terms with correct spellings]

Numbers: [e.g. numerals for 10 and above; currency format]

Markers: [inaudible], [unclear: ...], [crosstalk], [laughter]

Removed in clean verbatim: [fillers to remove; words to keep]

Confidential details: [what to redact and how]

---

Worked example

This example is fictional.

Raw output:

"ok so um for the q4 budget we're we're not going to approve the the extra 15 thousand jordan can you send the revised figures by friday yeah I can do that by friday"

After cleanup (clean verbatim, checked against audio):

[00:22:10] PRIYA: OK, so for the Q4 budget, we're not going to approve the extra fifty thousand. Jordan, can you send the revised figures by Friday?

[00:22:18] JORDAN: Yeah, I can do that by Friday.

What changed: fillers and repeats were removed; the listening pass corrected "15 thousand" to "fifty thousand"; the single block of text was split between two speakers; and timestamps were added before the decision. The "not" was checked carefully, because without it the decision reverses.

Find and replace, carefully

Find and replace is useful for a name the model consistently gets wrong, but check each match rather than replacing all at once. Replacing "Jon" with "John" across the file also changes "Jonathan" and "Jonas".

Should you use AI to clean up a transcript?

An AI model can tidy punctuation and remove fillers quickly, but it can also smooth over errors, change wording or add sentences that sound plausible. If you use one:

• Compare its output with the original, especially names, numbers and negatives.
• Do not ask it to fill passages you could not hear. A text-only cleanup model cannot verify missing speech from the recording.
• Check where the text goes. Pasting a transcript into an online AI service sends it to that provider. Only submit text you are allowed to share.

For a saved Meetings Brief live-notes result, choose Edit transcript, make your corrections, then Save and update brief. Saving requests a new brief from the corrected text, which goes to Anthropic's AI; daily limits apply. Earlier emails remain unchanged, and emailing the corrected version is a separate choice. See the privacy policy.

For a file transcript, download the text and edit your copy in a text editor. If you are fixing subtitles, change the SRT or VTT file too; a corrected text document does not update a previously downloaded subtitle file. The subtitle guide covers cue structure and timing review.

Frequently asked questions

How long does transcript cleanup take?

It depends on audio quality, the speakers and the required standard. Allow time to listen through the source, resolve uncertain words and review meaningful edits. Processing time and the amount of correction vary; a fluent draft still needs checking.

Can I clean up a transcript without the audio?

You can fix formatting and obvious typos, but you can't confirm names, numbers or misheard words. Mark the transcript as unchecked if the audio isn't available.

Should I correct grammar in a quote?

Follow the agreed quotation and transcript policy. Removing a filler is different from rewriting grammar or removing a qualification. Check the quote against the audio, and label a paraphrase as your account rather than putting it in quotation marks.

Why does the transcript include words nobody said?

Speech models sometimes produce text during silence, music or noise, or repeat a phrase. Check sentences in quiet passages against the audio and delete anything that wasn't spoken.

What is the most important thing to check?

Anything someone will act on: names, numbers, dates, decisions and who agreed to do what.

Sources and scope

Product details, W3C transcription guidance and the linked Whisper research checked on 4 November 2026. The four-pass method, style sheet, correction log and fictional examples are original working aids. Follow your project's record and retention standards when they require a different level of detail.