Verbatim vs Clean Verbatim Transcription

Full verbatim and clean verbatim describe how much of a recording a transcript keeps. Full verbatim writes down every word and the sounds that go with it, including filler words, false starts and stutters. Clean verbatim keeps every word that carries meaning and drops the noise. The usual choice for business, media and meeting work is clean verbatim, and legal records and conversation research usually need full verbatim. The style names are not standard, so the choice belongs in a written order before anyone transcribes.

Full verbatim keeps every sound; clean verbatim keeps every meaningful word

Full verbatim is the strict record. Rev defines a verbatim transcript as one where the transcriptionist types "all of the words they hear," including filler words, false starts, self-corrections and stutters ¹. The rule is about the audio, not the reader: if a speaker said "um," the transcript shows it.

Clean verbatim keeps the meaning and removes the clutter. GoTranscript, a transcription company, defines it as keeping "the meaning but removes obvious filler words, stumbles, and minor repetition" ². The wording itself stays as spoken, which is what separates it from the edited styles described below.

Removing fillers is usually safe for reading. Readability studies cited by Lease and Johnson in 2006 found that fillers and speech repairs may be deleted from transcripts without compromising meaning ³. Fillers are not always empty, though. Clark and Fox Tree proposed that speakers use "uh" and "um" to announce a delay, a minor one after "uh" and a major one after "um" ⁴. A clean transcript removes that signal, which matters if the delay is part of what you study.

The table shows what each style does with the main features of speech.

Intelligent verbatim rewrites for clarity, and its labels vary between providers

Intelligent verbatim, which GoTranscript calls edited transcription, goes one step further. GoTranscript describes its edited level this way: it "rewrites for clarity and flow, which helps readability but moves further away from the exact spoken record" ⁵.

The labels are the weakest part of the scale. Oliver, Serovich and Mason describe denaturalized transcription as removing "idiosyncratic elements of speech," such as stutters, pauses and nonverbals ⁶. The same authors suggest that transcription is "a powerful act of representation," which means the editing choices shape what the reader learns ⁷.

Practitioners disagree about how far to edit. The Minnesota Historical Society's oral history guide describes one view that a transcript should read "as smoothly as possible," with almost cut-and-paste editing ⁸. It also notes that each editor edits somewhat differently, so text one editor thinks is unimportant may be "extremely valuable" to a researcher ⁹. A provider's "intelligent verbatim" may therefore differ from another's, and the rules have to be written down. The next section shows the same passage in three styles.

The same passage in three styles shows exactly what changes

The example below was written for this guide to show the mechanics. It is not a transcript of a real recording. Suppose the audio says:

"Um, so we, we were going to, uh, ship it in March. No, sorry, April. The, the vendor pushed it back."

Clean verbatim removes the fillers, the repeated "we" and "the," and the false start "in March," but it keeps the speaker's wording. Intelligent verbatim also rewrites the sentence so it reads as writing. Both changes are legitimate choices, and both change the record, which is why the order should name the style.

Rev's own definitions draw the same line. Verbatim covers "all of the words they hear," including filler words and stutters ¹, while its non-verbatim style is "lightly edited by the transcriptionist for readability" ¹⁰.

Legal records, conversation analysis and some research need full verbatim

When a court record is the product, the rule is written into law. For federal courts, 28 U.S.C. 753(b) says each session "shall be recorded verbatim by shorthand, mechanical means, electronic sound recording, or any other method" ¹¹. Under the same statute, a transcript of a federal court proceeding does not count as official unless it is made from records certified by the reporter ¹². Court and client formats differ, so ask for the required format in the order.

Conversation research treats disfluency as data. In a corpus of task-oriented conversations, disfluency rates were higher when speakers discussed abstract figures, which the authors read as an increase in planning difficulty ¹³. A clean transcript cannot answer a question about hesitation, because it has removed the hesitation.

Practice differs by field. In nursing research, verbatim transcription of interview data is "widely considered to be integral" to analysis ¹⁴. Choose full verbatim when the way something was said is part of the evidence, and clean verbatim when only the content will be coded.

Most business, media and meeting transcripts use clean verbatim

Rev describes its non-verbatim style as "the most common" type of transcript it produces ¹⁰. That is one provider's statement about its own work, but it fits the reason most teams choose clean. It gives a faithful record of what was said, in text people will actually read.

Full verbatim costs more to produce. Rev says it "requires a ton of extra time and attention to detail," which is why it "costs a little bit extra" ¹⁵. By some estimates, one hour of audio or video can take 4 to 9 hours to transcribe, depending on the subject, the number of speakers and the audio quality ¹⁶. Every extra sound a transcriber must record adds to that work.

For most meetings, clean verbatim is the practical default. Our guide to meeting transcription covers the tools and the record-keeping, and the guide to how to transcribe an interview covers research interviews. If you only need to try a recording, the free audio to text tool in the browser produces a draft you can compare against both styles.

Style guides set the conventions: tags, inaudible marks, numbers and spelling

Two providers can call their style the same thing and still disagree on the details. The conventions below decide how a reader sees a transcript, so they belong in the order too.

The first decision is how to mark speech nobody can make out. GoTranscript says to use " [inaudible mm:ss]" when you cannot understand what is said ¹⁷, and " [crosstalk mm:ss]" when people talk at once and you cannot separate their words ¹⁸. The Minnesota guide uses a different marker: when a word is undecipherable, "leave a space and type [unclear]" ¹⁹. Pick one set and use it everywhere. Our question on marking inaudible speech works through the cases.

The second decision is what to do with sounds. GoTranscript says "background sounds or important noises should be noted," and to insert them "in square brackets" ²⁰. The third is spelling: "choose either American or British English spelling, and keep it uniform across the transcript" ²¹.

For false starts, the Minnesota guide is specific: "If the narrator begins with several words and then reorganizes his or her thoughts and resumes, delete the false start" ²². That rule suits a publication. A legal or research transcript would keep the false start, so the rule belongs to the style you chose. The same guide recommends a two-stage workflow: "the initial draft of the transcript should be a nearly verbatim copy of the recorded interview," and editing comes later ²³. That is a good model even for clean output, because the editor can always check the draft against the audio.

AI transcribers can drop fillers, which you may need to undo

Some automatic transcribers drop fillers, which can break a verbatim requirement without anyone noticing. Whisper is one example. A 2025 study on dementia speech reports that "Whisper tends to omit filler words such as uh and um in its transcriptions" ²⁴. The study is a preprint and its finding is a side observation, so treat it as a warning to check, not a measured rate.

A written instruction may not fix this. OpenAI's Whisper prompting guide says that Whisper "follows the style of the prompt, rather than any instructions contained within" ²⁵. This suggests that typing "keep the fillers" into the prompt may not work. Instead the prompt has to look like the output you want. This is an inference from the guide's behavior description, and a test on your own audio is the only way to confirm it for a given tool.

Tools differ. A 2025 study of non-native English speech found "significant variation in performance across systems and disfluency types" when recognizers handled filler words, repetitions and revisions ²⁶. Test the tool on recordings like yours, with the disfluencies you need.

Researchers who need hesitations have started to label them. Lin and colleagues note that "many ASR systems discard or generalize hesitations, losing important acoustic details" ²⁷. In their 2025 challenge on non-native English, a fine-tuned Whisper system scored "5.81% WER (Extra)" when it labeled hesitations, against "6.47% WER (Pure)" when it removed them ²⁸. The chart also shows Whisper Large V3 Turbo, where fine-tuning on labeled filled pauses gave 5.5 percent against 6.2 percent when hesitations were removed ²⁹. That is a gap of 0.7 percentage points ³⁰. The paper compares annotation schemes for fine-tuning, so these scores show which labeling a verbatim system needs, not which output a clean reader would prefer.

Machines and people also make different filler errors. In a 2017 comparison on telephone speech, the recognizer "tends to confuse filled pauses ("uh") and backchannel acknowledgments ("uhhuh")," and "Humans tend not to make this error" ³¹. If a draft has many short "uh" and "uhhuh" passages, check those first.

Human drafts also need checking. Zayats and colleagues report that a quality checking pass improved the average transcription error rate by 5 to 20 percent ³². Human transcribers in a 1997 review of spontaneous telephone conversation had an error rate of 4 percent ³³, which is a useful floor when you judge a draft. Word error rate is always measured against a reference transcript. Researchers describe it as "a standard measure of discrepancy between machine and human transcriptions" ³⁴. If the reference drops fillers, a verbatim draft will score differently from the same draft measured against a verbatim reference. Our glossary entry on word error rate explains the formula, and the question on transcription QA covers how to proofread a draft against the audio.

Agree these rules before anyone transcribes

A short written order removes most of the disagreement. Agree these points with the people who will transcribe, edit and use the file:

1. Style and level: full, clean or edited, named in one sentence.
2. Fillers and false starts: kept or removed, and whether repeated words stay.
3. Stutters and repetitions: kept as spoken or cleaned up.
4. Grammar: left as spoken or corrected.
5. Sounds and non-speech: which ones to note, and in what brackets.
6. Tags: one set for how to mark inaudible speech, [crosstalk] and unclear speech, with timestamps.
7. Spelling and numbers: one spelling variety, and a rule for writing numbers.
8. Speaker labels and timestamps: whether each speaker change gets a name and a time, covered in our guide to timestamps and speaker labels and the speaker diarization entry.
9. Review: who checks the draft against the audio, and how many passes it gets.

Give the transcriber a sample of two or three minutes, transcribed to the rule, before the full job. A sample exposes the disagreements that a style name hides. For a concrete look at a finished transcript in each style, see the question on what a good transcript looks like. Our overview of transcribing audio to text shows where a human or an AI draft fits in the workflow, and human vs AI transcription compares their accuracy and cost. For the wider subject, see our transcription overview.