the short answer
write [inaudible 00:14:32]: square brackets, the word, and the timestamp where the gap starts. the timestamp is the important part. without it, nobody can check your call.
that's the convention most commercial transcription services follow, and it works for journalism and research too. the rest of this page covers the cases where a different tag is more honest.
inaudible vs. unintelligible
they sound interchangeable. they describe different problems.
[inaudible]— you can't hear it. the speaker turned away from the mic, a truck went by, the recording dropped out.[unintelligible]— you can hear it, but you can't make out the words. mumbling, a heavy accent you don't know, someone speaking with their mouth full.
the difference tells the reader whether a better listener might recover the words. unintelligible audio sometimes yields to a native speaker or a second pass with headphones. inaudible audio usually doesn't. court transcribers use indiscernible for the second case.
the common tags
| tag | use it when |
|---|---|
[inaudible 00:00:00] | speech is there but can't be heard |
[unintelligible 00:00:00] | speech is heard but the words can't be made out |
[crosstalk 00:00:00] | two or more people talk over each other and neither is recoverable |
[word?] or [unclear: word] | you have a best guess but aren't sure |
[pause] | a silence long enough to matter. some guides only tag pauses over 10 seconds |
[laughs], [laughter] | one speaker laughs, or the room does |
[background noise], [music] | a non-speech sound that explains a gap or changes the meaning |
[audio cuts out] | the recording itself has a gap |
rules that hold across every style
- timestamp every uncertainty tag.
[inaudible]on its own is a dead end.[inaudible 00:14:32]is an instruction: go listen here. commercial services typically treat an untimestamped inaudible tag as an error. - don't tag what you could have heard. the tag is for audio that defeats a careful listen, not for a first pass you didn't want to replay. overusing it is the most common transcription mistake.
- never fill the gap with a guess. if you think you know the word, mark the guess as a guess. a silent guess becomes a quote, and a misquote becomes a correction.
- keep the tag set small. pick your tags once, write them down, and stick to them. the reader shouldn't have to decode a new tag on every page.
- square brackets, not parentheses. brackets mark the transcriber's words. parentheses get confused with the speaker's own asides, and in some conventions they mean something else entirely.
by field
journalism
for reporters, an inaudible passage in a quote is a stop sign. you can't quote around it with ellipses as if the speaker paused, and you can't fill it. go back to the audio, and if it's still unrecoverable, paraphrase or ask the source. timestamps matter most here: your editor or fact-checker will want to hear the passage themselves.
qualitative research
there's no single academic standard. most projects use bracketed tags like [inaudible], [unclear], and [laughs], and define them in a transcription key in the methods appendix. conversation analysis is the exception: it uses jefferson notation, where ( ) marks talk that couldn't be heard and (word) marks a best guess.
legal transcripts
court transcripts use parentheticals, and the usual words are (inaudible) and (indiscernible). a live court reporter isn't supposed to need them at all: California's Court Reporters Board tells reporters to interrupt the proceeding rather than lose the record, and reserves "inaudible" for transcripts made from audio recordings. if you're preparing a transcript for a court, check the local rules. some courts specify exactly which words are allowed.
tags in AI transcripts
speech models don't write [inaudible]. they write their best guess, and they write it with the same confidence as everything else. that's the dangerous part: a clean-looking sentence that was never said.
a usable AI transcript flags low-confidence words so you know where to listen. ours underlines them, and every word links to its second of audio, so checking a doubtful passage is one click. when the audio really can't be recovered, you replace the guess with a tag and move on.