How to Transcribe an Interview for Editing

Transcribe an interview for editing by running it through an AI speech-to-text engine first, then correcting the output against the audio while the transcript keeps word-level timecode and speaker labels, so every line you later highlight maps to an exact, trimmable range on the recording. Skip any one of those three things, timecode, speaker labels, accuracy you've actually checked, and you've produced a document to read, not a tool to cut from. Here is the full process, what it costs right now, and where people quietly ruin an otherwise good transcript before they ever open an NLE.

Editing transcripts vs reading transcripts

Why a transcript for a reader and a transcript for an editor are different objects

A transcript with paragraph-level or sentence-level timestamps is fine for a journalist scanning for quotes. It is not enough to build a cut, because there's no way to know exactly where a given word starts and ends inside that paragraph. Editing needs every word individually anchored to a frame of timecode, so that highlighting a sentence on the page selects a precise, trimmable range on the actual recording. Without that, the transcript and the footage stay two separate things you have to manually reconcile every time you want to cut something, which is exactly the busywork a transcript-based editor exists to remove.

This is also where speaker labels come in. Automatic speaker identification, or diarization, is what turns a two-person interview from one unbroken wall of text into a readable script with a name attached to every line. Most transcription tools now handle this by default, but diarization accuracy degrades with crosstalk the same way word accuracy does. Check the speaker boundaries during your correction pass, not just the words, especially anywhere subjects talk over each other.

How accurate AI transcription actually is

Clean audio and real interview audio are two different benchmarks

Vendors like to quote word error rates from clean, single-speaker benchmarks such as LibriSpeech, and modern large models do perform well there, often in the low single digits. That number does not survive contact with a real interview. On messier real-world audio, meetings, podcasts, phone calls, word error rates for the same class of models typically run 8 to 12 percent, and that's before you add crosstalk, an accent the model wasn't trained on, a noisy location, or a cheap lav mic buried under a jacket.

Do the arithmetic on that: even a tight 5 percent error rate means roughly one wrong word every twenty, and errors don't distribute evenly. They cluster on names, numbers, and technical jargon, which happen to be the words most likely to break a sentence's meaning if nobody catches the swap. The practical rule we'd give any editor: AI transcription is the right default for speed and cost, but budget an actual correction pass on anything that isn't a clean, single-speaker studio recording. A multi-speaker interview with any crosstalk needs a human read-through before you trust the text enough to select from it.

In ScriptCut, lines tagged Crew never get pulled into any AI selection tool, so slate, set talk, or interrupted crosstalk can stay in the transcript without polluting your selects later.

Seeing the gap between a plain transcript and one actually built to drive an edit makes the distinction concrete.

What transcription costs in 2026

AI pricing, as of October 2026

Prices and plan names shift often enough that it's worth checking the vendor page before you commit a season's budget, but here's where things stood this month.

Tool	Price	Notes
Rev	$0.25/min for AI transcription, $1.99/min for human transcription	Free tier includes roughly 45 minutes of AI transcription a month
Otter.ai	Free Basic (300 min/mo), Pro around $8.33/mo billed annually	Business plan removes the per-meeting cap at a higher per-seat rate
Descript	Free tier available, paid tiers from roughly $12 to $20/mo billed annually	Ties transcription to its own timeline editor rather than a separate pre-edit step

Human transcription still exists for a reason: heavy accents, dense technical jargon, legal or medical terminology, and archival audio with real degradation are all places a trained human still beats a model. It costs more and takes longer, but it's the right call once the source material pushes AI's error rate past what a correction pass can reasonably fix.

Where this method comes from

AI didn't invent transcript-driven editing, it just made it cheap

Editing from a written transcript predates any of the current software by decades. Michael Rabiger's Directing the Documentary (Focal Press, first published 1987) formalized the idea of building a cut from a logged, timecoded transcript of what subjects actually said, rather than from a pre-written script, long before AI speech recognition existed. Before that, editors did the equivalent work by hand: typed transcripts cut apart with scissors and taped back together in order, and radio producers at NPR physically splicing quarter-inch tape with a razor blade to arrange a radio cut as early as the 1970s. What's changed is cost and speed, not the underlying logic. A timecoded transcript used to take a logger hours to build by hand; now a machine does the first pass in roughly real time, and text-based editing has moved from a niche documentary technique into mainstream NLEs, Adobe added its own version to Premiere Pro at NAB 2023.

That history matters for a practical reason: the core discipline, verify every selected line against the actual recording before you lock anything, was true when editors were cutting paper with scissors and it's still true now that a model is doing the first pass. The tool got faster. The need to check its work didn't go away.

Correcting the transcript before you cut

A read-through pass is not optional for multi-speaker audio

The correction step is where most AI-first workflows quietly fall apart, usually because nobody budgets time for it. Play the transcript against the audio at normal speed, not 2x, and fix three things as you go: misheard words (especially proper nouns, jargon, and numbers), missed or wrong speaker labels, and any place the model inserted a word that was never said, which is rarer than a dropped word but far more dangerous. If two people talk over each other, decide now whether you need both lines legible or whether one subject's words can be marked as the throwaway.

How long this takes depends entirely on source quality. A clean two-mic podcast interview might need five minutes of spot-checking per thirty minutes of audio. A phone interview recorded over a bad connection, or a documentary subject with a strong regional accent the model hasn't seen much of, can eat as much time as a from-scratch human transcript would have cost. Know which kind of audio you're dealing with before you promise a turnaround time to a client or a producer.

From corrected transcript to rough cut

What actually happens once the text is trustworthy

Once you have a timecoded, speaker-labeled, corrected transcript, the workflow is straightforward: read it, highlight the moments worth keeping, trim filler inside the lines you keep, and arrange the survivors into a paper edit that holds together as a story. This is the same method BBC's factual series Empire used on a three-camera interview, paper-editing it in text before conforming the picture in DaVinci Resolve, a sequencing step editor Dilesh Korya has said saves real days of timeline work compared to cutting cold.

In ScriptCut's text-based editing flow, every highlighted selection inherits its exact in and out point from the word-level timecode underneath it, so there's no manual reconciliation between the page and the footage. The finished sequence then hands off to your NLE: ScriptCut writes out an editable XML timeline your edit software can open directly, alongside a clean subtitle file and the audio itself, so DaVinci Resolve, Premiere Pro, Final Cut Pro, and Avid all pick up the same cut without a manual re-sync. For projects with several source recordings, ScriptCut builds a separate timeline for each file, which keeps a multi-camera or multi-session interview organized before it ever reaches your timeline. If you later need the transcript itself as a clean caption file, our free SRT to VTT converter handles the format conversion without touching your edit.

For a sense of real pace: a 40-minute two-person interview with decent audio can realistically go from raw recording to a locked rough cut in under an hour. A couple of minutes for the AI pass, some correction time depending on audio quality, then selection and arrangement against timecode that's already accurate. That math changes fast if the audio is rough, which is exactly why the correction step above isn't something to skip to save time.

Where this goes wrong

Five mistakes that show up constantly in transcripts built for editing

• Transcribing without word-level timecode. You end up with a document to read, not a tool to cut from, and you'll be scrubbing the raw file by eye to find the exact frame anyway.
• Trusting the first AI pass blindly. Even a strong 95 percent accuracy rate means an error roughly every twenty words, and errors cluster around names, numbers, and technical terms, exactly the words that matter most if you're quoting someone on record.
• Ignoring the recording itself. No transcription engine fixes bad audio after the fact. A better mic and a quiet room do more for your accuracy than any model setting ever will.
• Skipping the speaker-label check. Diarization errors are easy to miss on a quick skim and expensive to discover after you've built a sequence around the wrong attribution.
• Treating the transcript as the finished product. It's a working surface for editing decisions, not an archive. Words on a page can't show you tone, pacing, or delivery, so a final listen to your selected clips still matters before anything locks. Documentary editor Errol Morris has gone on record refusing to work from paper cuts at all precisely because performance doesn't survive the trip to text, a reasonable caution worth remembering even if you don't share his conclusion.

The honest tradeoff

Speed costs you some certainty, and that's fine if you plan for it

AI transcription trades a small, real amount of accuracy for a large amount of speed and cost savings, and on clean audio that trade is easy to make. On messy audio, the correction time can eat past what human transcription would have cost in the first place, so be honest with yourself about your source quality before you commit to an all-AI pipeline on a project where getting a quote wrong has real consequences. Either way, a transcript is a means to an edit, not an end in itself. Get the words right, check the tone on the clip, and the cut follows.

If you want to see this whole workflow end to end rather than piece by piece, our guide on editing an interview faster with a transcript-first process walks through the full sequence from raw audio to export.

Ready to turn a corrected transcript into a timed sequence instead of a flat document? Open your recording in ScriptCut and build the selects while the transcript is still fresh.

Sources

• Gladia: Whisper vs Google Speech-to-Text vs Amazon Transcribe
• Rev: Pricing
• Otter.ai: Pricing
• Descript: Pricing