What Is Verbatim Transcription?

To make the differences concrete, here is one short example exchange written in all three styles:

The example: A speaker pauses, restarts, and answers a question.

Strict verbatim

Interviewer: So, um, can you tell me about, uh,
the project?

Subject: Yeah, yeah, sure. So I, I started — well,
we started, um — [coughs] excuse me — we started
the project back in, um, January? Yeah, January.
And it's been, you know, it's been a journey.
[laughs] Yeah.

Intelligent verbatim

Interviewer: Can you tell me about the project?

Subject: Sure. We started the project back in
January. It's been a journey.

Clean read

Interviewer: Tell me about the project.

Subject: We launched the project in January, and
it has been an interesting journey since then.

All three are accurate to what was said — they just preserve different amounts of how it was said. Choose the style based on what you'll do with the transcript.

Strict verbatim is required (or strongly preferred) in these contexts:

• Court reporting and legal proceedings. Depositions, trial transcripts, and other legal records must capture exactly what was said. Filler words, hesitations, and tone of voice may matter to a case. In US federal courts, the law requires proceedings to be recorded verbatim (28 U.S.C. § 753(b)).
• Qualitative research (PhDs, sociology, ethnography). When studying how people think and speak, the filler words, hesitations, and self-corrections are data. Researchers analysing speech patterns need every utterance.
• Medical transcription. Specifically, mental health and behavioural assessments where the manner of speech is part of the diagnosis.
• Linguistics and conversation analysis. Studying turn-taking, interruptions, repair sequences, and other conversational structures.
• Insurance investigations. Where exactly how a claimant phrased something can matter.
• Focus group research. Where the dynamics of how people respond — including filler and hesitation — are part of the analysis.
• Forensic linguistics. Authorship attribution, threat analysis, and stylometric work all depend on capturing every speech feature.

If you're writing a transcript for any of these purposes, request strict verbatim explicitly when ordering from a service or when configuring an AI transcription tool. Many AI tools drop filler words by default, so check the tool's settings.

Intelligent verbatim is the right choice for most general business and content use cases:

• Journalism interviews. Reporters need accurate quotes that preserve the speaker's voice without making them look inarticulate by including every "um."
• Podcast show notes. Capture what was said in readable form without forcing readers to wade through hesitations.
• Customer feedback analysis. Understand what customers said in research interviews without getting distracted by speech artifacts.
• Sales call review. Reps want to study what they said and how prospects responded — readability matters more than capturing every "right" and "yeah."
• Conference talks and keynotes. Distributing a written version of a talk for accessibility or content reuse.
• Internal meeting notes. Capturing decisions and discussion without the noise of speech disfluencies.
• Educational content. Lectures and tutorials transcribed for student reference.

Many AI transcription tools produce something close to this by default. Whisper often leaves out filler words, while Rev's API keeps "um" and "uh" unless you turn on its remove_disfluencies option. On TranscribeVideo.ai, a YouTube transcript is the video's own captions, so its style matches those captions. TikTok and Instagram transcripts come from speech-to-text on the audio, so check them against the video if you need every filler word.

Clean read transcription is appropriate when the final output will be read as prose, not as a record of speech:

• Repurposing audio into blog posts. Convert a podcast episode into a written article — clean read removes spoken-language patterns that don't read well.
• Book content from interviews. Authors who interview subjects often want clean-read transcripts they can edit further into prose.
• Marketing case studies. Customer interview content turned into branded marketing material reads more polished in clean read style.
• Email newsletter content from podcasts. The conversational style of audio doesn't translate to skimmable email writing.
• Academic papers citing interviews. When the meaning matters more than the exact phrasing, clean read makes citations more readable.

Clean read is rarely the output of an AI transcription tool directly — it usually requires a second pass through an editor (human or AI) that paraphrases and restructures. Tools like Descript can remove filler words, which gets part of the way there.

By default, modern AI transcription tools clean up speech disfluencies. To get strict verbatim, you have to request it explicitly:

Whisper (OpenAI)

Use the prompt parameter: --initial_prompt "Umm, so, uh, I was, like, thinking... hmm." in the CLI, or the prompt parameter when calling the API. Whisper often leaves out filler words. OpenAI's guide lists keeping filler words as one use for the prompt, and notes that Whisper does not follow instructions like a chat model. So write a prompt that itself contains fillers. The prompt works as an example of the style you want, not as an instruction. It nudges the output toward strict verbatim, but it does not guarantee that every filler is kept.

Otter.ai

We could not find a strict verbatim mode in Otter's documentation. Test a sample transcript to see how it handles filler words before you rely on it.

Rev.com

Rev's human transcription service offers a verbatim add-on that captures every word, filler and nonverbal sound. Rev's speech-to-text API also has a verbatim option that keeps false starts and fillers.

Trint

We could not find a filler-word setting in Trint's help center, so check a sample transcript before you rely on it for strict verbatim.

The honest tradeoff

AI transcription is much weaker at strict verbatim than at intelligent verbatim. Many speech models tend to drop fillers, false starts and repeats, even when you ask them to keep them. For high-stakes strict verbatim work (legal, research), human transcription is still the standard. AI is a good fit for intelligent verbatim, with a human check for names and terms, and a useful first draft for clean read.

Beyond the words, verbatim transcripts have formatting conventions:

• Speaker IDs in caps with colon. JOHN: or INTERVIEWER:. Some research conventions use [P1], [P2] for participants or pseudonyms.
• Bracketed non-speech sounds. [laughter], [cough], [pause - 3 sec], [phone ringing], [overlapping speech].
• Filler words written out. "um," "uh," "ah," "er," "mm-hmm," "uh-huh." Convention varies on apostrophes.
• False starts marked with em-dash. "I was going to — well, actually, let me start over."
• Pauses marked. Short pauses with ellipsis or (...). Long pauses with a duration: [pause - 5 sec].
• Inaudible markings. [inaudible] or [inaudible 00:23:14] when a section can't be transcribed.
• Overlapping speech. [crosstalk] or square brackets around the overlapping sections to show simultaneity.
• Timestamps. Inserted at speaker changes or at fixed intervals (every 30 seconds, every 2 minutes) for navigation. Format: [00:14:25].

For research transcripts especially, follow the conventions used by your field's standards. Sociolinguistics has different conventions from clinical interviewing.

Is verbatim transcription required by law?

For court records: often yes. US federal law, for example, requires court proceedings to be recorded verbatim (28 U.S.C. § 753(b)), though how fillers are handled can vary. For accessibility captions: no. W3C's WCAG guidance asks captions to show the dialogue, who is speaking, and meaningful sounds, but it does not require strict verbatim. For research: it depends on the study's method, not on the law.

How accurate is AI verbatim transcription?

There is no single figure. It depends on the audio and the tool. For one published reference point, OpenAI's Whisper paper (Radford et al., 2022) reports a 2.7% word error rate (roughly, the share of words it got wrong) for Whisper large-v2 on clean read audiobooks (the LibriSpeech test-clean benchmark), but 16.9% on recorded meetings with headset mics (AMI IHM). Those benchmark scores are counted after OpenAI's text normalizer, which removes differences that don't change the meaning, so they say little about strict verbatim. For strict verbatim — every filler and false start — expect AI to drop some of them as it goes, and plan a human pass. Human services remain the standard here; Rev, for example, says its human transcripts are 99%+ accurate (its own figure).

How long does verbatim transcription take?

Human transcription takes several times the length of the audio, and strict verbatim takes longer than intelligent verbatim. AI takes minutes per hour of audio, regardless of style. The choice between human and AI usually comes down to whether the use case can tolerate AI-level accuracy or requires the higher human standard.

Can I convert AI intelligent verbatim back to strict verbatim?

Not reliably. AI cleaning is lossy — once "um" and "uh" are removed, they can't be reconstructed without re-running the transcription on the original audio with explicit instructions to preserve them. If you anticipate needing strict verbatim, configure for it from the start.