How to edit a transcript fast: seven mistakes to fix before you publish
How to edit a transcript fast: the seven mistakes speech recognition makes most, the quickest search for each, the fix, and a time budget per hour of audio.
To edit a transcript fast, don’t read it top to bottom against the audio. Hunt for seven specific mistakes instead: wrong names, wrong numbers, homophones, phantom text over silence, missing punctuation, unmarked speaker changes, and clipped openings and endings. Each has a search that surfaces most cases in a minute or two, and the timestamps tell you which few seconds to listen to. Done this way, a full pass on an hour of two-person audio takes about 45 to 60 minutes.
The searches below assume you’ve copied the transcript into an editor with find and replace. The transcript view in FreeTranscribe has a search box, but it’s a plain, case-insensitive match, fine for one word and not for patterns. Keep the tool tab open beside the editor: every line on screen carries its start time, as do the SRT and VTT exports, so a flagged phrase leads straight to the right spot in the audio.
| Mistake | Fastest way to find it | Fix |
|---|---|---|
| Wrong names and proper nouns | Regex search for capitalised words, checked against a names list | Find and replace each wrong spelling, whole word only |
| Wrong numbers, dates and units | Search for digits and for fifteen, fifty, thirteen, thirty and the rest | Listen to each hit; apply one number style |
| Homophones and can versus can’t | Search for each word in the small confusable set | Fix on sight; listen for can and can’t |
| Phantom text over silence | Search “thank”, “subscribe”, “watching”; scan for repeated lines | Delete the run once you’ve checked the audio is empty |
| Missing punctuation | Regex for long runs without a full stop; search question words | Break sentences where the speaker breathed |
| Unmarked speaker changes | Search for turn-opening phrases | Add a bold speaker name and line break at each turn |
| Clipped opening and ending | Play the first and last ten seconds | Type what was said |
The time figures under each heading are planning numbers, not measurements. Clean audio with one speaker comes in well under them; a noisy panel with jargon does not.
Names and proper nouns are wrong more often than anything else
Speech recognition spells a name the way it sounds. Sarah Smyth becomes Sarah Smith, a small company becomes two common words, and a product name becomes whatever it rhymes with. The base model is worst at this, and the wrong name is the mistake that embarrasses you most.
Find them by searching for capitalised words. In an editor that supports regular expressions, \b[A-Z][a-z]+ with case sensitivity on steps through every capitalised word. Skip sentence openers and read the rest, including names the model split in two or turned into ordinary words.
Before you start, write a names list: everyone who speaks, everyone they mention, companies, products, places, titles. Get spellings from a bio or an email signature, not from memory. Then find and replace each wrong spelling in a single pass, with “match whole word” on so “Smith” inside “Smithsonian” stays put.
Rough planning figure: 5 to 10 minutes per hour of audio once the names list exists.
Numbers, dates and units get swapped for their neighbours
“Fifteen” and “fifty” sound alike, and so do “thirteen” and “thirty”. Currencies vanish (“two hundred” with no pound or dollar), dates lose their year, and “per cent” comes out three different ways. A wrong figure is the mistake most likely to be quoted.
Search for digits with [0-9], and separately for the number words that get confused: fifteen, fifty, thirteen, thirty, fourteen, forty, sixteen, sixty, seventeen, seventy, eighteen, eighty, nineteen, ninety. Listen to each hit. It’s the only reliable check, because the transcript reads fine either way.
While you’re there, pick one style and apply it: digits or words, a currency sign before or the word after, “20 per cent” or “20%”. Consistency is a find-and-replace job once the numbers are right.
Rough planning figure: about 5 minutes per hour of audio, longer for a finance or research conversation.
Homophones and near-misses read correctly and mean the opposite
These are the mistakes a spellchecker can’t see. “Their” for “there”, “to” for “too”, “accept” for “except”, and the dangerous one, “can” for “can’t”, where a dropped consonant flips the sentence. “We can ship in March” and “We can’t ship in March” are one soft letter apart in speech.
Search for each word in the small set that matters: “can”, “can’t”, “won’t”, “want”, “their”, “there”, “they’re”, “your”, “you’re”, “then”, “than”, “affect”, “effect”. Read each hit in its sentence. Most are right, and the wrong ones are obvious from context. For “can” and “can’t”, listen. Context doesn’t help when both readings make sense, and that’s exactly when the error costs you.
Rough planning figure: about 10 minutes per hour of audio. Most hits pass on sight.
Phantom text appears over silence and repeats itself
Where the audio goes quiet, at the start before anyone speaks, over a long pause, or under a music bed, the model can write something anyway. It’s typically a sign-off such as “Thanks for watching”, a phrase copied from nearby speech, or one short line repeated several times. The post on why speech recognition invents words in silence explains the cause and how to trim silences beforehand.
Finding it after the fact is quick. Search for “thank”, “subscribe” and “watching”, and check any hit that doesn’t fit the conversation. Then scan for repeats: in the tool’s transcript view, a repeated phrase shows as the same words on consecutive lines. Listen to the first one; if the audio is empty, delete the whole run. Also check any line where the timestamp jumps by several seconds with little text between, because that’s a pause the model may have filled.
Rough planning figure: about 5 minutes per hour of audio.
Missing punctuation turns speech into run-on sentences
Speech recognition punctuates on its own, but it treats a conversational ramble as one sentence more often than a reader can bear. Questions lose their marks and a two-minute answer arrives as one paragraph.
Search for long stretches without a full stop. In a regex editor, [^.?!]{300,} finds any run of 300 characters without sentence punctuation; adjust the number to taste. Read each hit and break it where the speaker breathed. Then search the question words that usually need a mark: “why”, “how”, “what”, “do you”, “did you”. The TXT export already breaks paragraphs about once a minute at a sentence end; add more at each change of thought.
If the transcript is going out as subtitles, punctuation matters more, because a cue with no sentence end feels unfinished on screen. How to make an SRT file covers the line and timing rules the export follows.
Rough planning figure: 10 to 15 minutes per hour of audio for a conversation, less for a prepared talk.
To edit a transcript with two voices, add the speaker changes yourself
The tool doesn’t label speakers. The transcript comes out as one continuous voice, and a two-person interview reads like one person arguing with themselves. If more than one person spoke, add labels by hand at each change. Do it before the punctuation pass, because a speaker change is also a paragraph break.
Use the turns you can see in the text. Search for the phrases that open one: “so”, “yeah”, “well”, “right”, “I think”, and the interviewer’s stock questions. Each hit is a likely change. Insert the speaker’s name in bold and a line break, then listen only where you’re unsure who’s talking. With two voices you can mark most changes on sight; a panel of four needs far more listening.
Rough planning figure: 10 to 15 minutes per hour of audio for two speakers. Double it for a panel.
The first and last ten seconds of every file often lose words
Models tend to clip the opening: a name spoken in the first second, the “Hi, I’m” before anyone has settled, a greeting under intro music. Endings suffer the same way, with the sign-off trailing into silence and getting cut or replaced by phantom text. You hear these parts least while editing, so they’re the mistakes that survive to publication.
There’s no search for this one, just a habit: play the first ten seconds and the last ten seconds of every file and read along. Twenty seconds per file catches a missing opening name more reliably than any pattern. If a recording opens on speech, add a second of silence at the front before transcribing so nothing lands on the very first frame.
Rough planning figure: one minute per file, regardless of length.
The order of the seven matters a little. Do names first, because one replace fixes every later occurrence. Do speaker labels before punctuation, because the labels create the paragraphs. Do the ten-second check last, once you know the voices.
If you’re starting from audio rather than a transcript, transcribe it in your browser first. It’s free, nothing is uploaded, and there’s no length limit. It runs in desktop Chrome or Edge, in English only, on the base model, taking about as long as the recording on a thin laptop, so the pass above is what makes its output fit to publish.
Frequently asked questions
Should I fix mistakes in the SRT file or in the text? Fix the TXT export, since that’s the copy people read. If you also need subtitles, run the same find-and-replace passes on the SRT. Line-by-line rewrites are easier in plain text, then pasted into a subtitle editor.
Do I have to listen to the whole recording? Not for a clean, single-speaker file; the searches and timestamps send you to the few seconds that need ears. Listen to the whole thing when the audio was noisy, the topic is technical, or a wrong word could cost someone money.
How do I find a name the model spelled three different ways? Search for the part it got right. If “Smyth” came out as Smith, Smithe and Smyth, searching “Smi” and “Smy” finds all three. Replace each variant separately with whole-word matching on.
What about “um”, “you know” and false starts? That’s a style decision, not a mistake. For a readable transcript, cut them. For a legal or research transcript, keep them. Decide once, apply it everywhere with find and replace, and don’t change the meaning of a sentence while tidying it.
Sources, checked 15 September 2026
- No third-party sources were needed. The post is written from practice; the facts about the tool’s search and exports come from its code.