A 40-minute interview might give the edit three minutes of usable sound, and finding them is the slow part. Someone scrubs the timeline, someone else keeps notes in a doc, and the director remembers “a great line about the factory, somewhere around minute twenty”.
A transcript turns all of that into text you can search and mark up, and later the same text becomes the subtitles.
Getting a transcript takes minutes now. Getting one you can show a client takes a bit more work, mostly in predictable places.
In short
- Use automatic transcription for internal work, and add a careful human pass before a client or viewer reads the text.
- Most errors come from the recording: overlapping voices, music under speech, uncommon names and a bad mic.
- Transcribe from the cleanest audio track and keep a glossary of names for each client.
- In the review pass, check names, numbers, brand spellings, jargon and speaker changes.
- Fix the transcript before making subtitles, since its errors carry into every subtitle file.
Three ways to transcribe a video
| Method | How long it takes | When it makes sense |
|---|---|---|
| Typing it yourself | Several times the running time | Short clips, audio the software can’t handle, material that can’t go to an outside service |
| Automatic | Minutes | Logging footage, finding a line, internal notes, rough cuts |
| Automatic plus a human pass | Minutes, plus one careful listen | Anything a client or an audience will read: subtitles, quotes, transcripts for approval |
Typing by hand still makes sense for a short testimonial with a heavy accent in a noisy room, which can be quicker to type than to correct, or when the contract keeps the footage in-house.
An automatic transcript is fine on its own when only the team will read it, to log an interview or check what was said in take four.
Once a client or a viewer is going to read the text, let the software do the typing and spend your time on the parts it gets wrong.
What trips up automatic transcription
Most errors trace back to the recording. Two people talking at once is the classic case. The transcript picks one voice, blends the two into a sentence nobody said, or drops the interviewer’s “right, right” into the middle of the answer.
Music under speech comes next. A bed that sounds fine in the mix competes with the voice, and the quieter words at the end of a sentence tend to go missing.
Then there’s vocabulary. The software writes the most common word that sounds like what it heard, so surnames, product names, acronyms and industry jargon come out plausible and wrong.
Accents and speakers who switch languages mid-sentence get the same kind of guess. If you know the language of the video, set it instead of relying on automatic detection, which a music intro or an opening line in another language can throw off.
A bad mic makes all of it worse. Camera audio from across the room or wind on an exterior shot costs accuracy, and no setting in the software gets it back.
Prepare the audio before you transcribe
Use the cleanest audio you have. If the interview was recorded with a lav or a boom on its own track, transcribe from that track and leave the camera’s scratch audio alone. If the cut already has music and effects, export a dialogue-only version for the transcript.
On set, ask each interviewee to say and spell their name and job title at the start of the first take. It takes ten seconds and saves a wrong lower third later.
Ask the interviewer to let each answer end before the next question, too, which helps the transcript as much as the edit.
Keep a glossary for each client, with the names of their people, products and acronyms written the way they write them. You’ll use it in every review pass and again on the next job.
Speaker labels and timestamps
Speaker labels come from diarization, the part of the software that tells voices apart. It doesn’t know who anyone is, so you get Speaker A and Speaker B.
Check the changeovers anyway. Short interjections land on the wrong person, and two similar voices can be merged into one. Replace the letters with names before the transcript leaves the edit.
A timestamp at the start of each paragraph is enough for finding things. Word-level timestamps are what let you click a word and land on it.
Either way, the times count from the start of the file. If your sequence starts at 01:00:00:00, or you transcribed an export that begins halfway through a clip, add that offset when you look for the moment in the timeline.
What to check in the review pass
Play the video with the transcript in front of you and fix whatever doesn’t match what you hear. Leave the grammar alone unless you’ve decided to clean it up (see below).
Go through it in this order:
- Names of people, companies and places, checked against the glossary or the client’s own material.
- Numbers, because a wrong figure in a subtitle is the kind of mistake a client remembers. Check prices, percentages and dates, and listen closely for pairs like fifteen and fifty.
- Brand and product names, which need the client’s exact spelling and capitalization.
- Technical terms and acronyms, which are where the software guesses most.
- Who’s speaking at each change.
If the tool shows a confidence score for each word, start with the lowest ones.
Before the pass, decide between verbatim and clean verbatim. For subtitles and quotes, most teams drop the ums, false starts and repeated words. A transcript for legal review or research keeps all of it.
Pick one per project so two people don’t edit the same interview two different ways.
What to do with the transcript
Subtitles come from the transcript, so fix the transcript first. A misspelled name there turns into a misspelled subtitle in the SRT, the VTT and the burned-in version. For line length, reading speed and timing, see our subtitle guidelines.
Finding a quote stops being a scrubbing job. Search the transcript for the word you remember, “factory” or “weekend”, and jump to it.
For a paper edit, read all the interviews, mark the best lines and copy them into a document in the order the story needs, each with its file and timecode:
INT_Ana_01 · 17:42 · “We moved the whole line in one weekend.”
INT_Carlos_02 · 04:18 · “The first Monday was chaos. By Wednesday it ran.”
INT_Ana_01 · 31:05 · “Nobody on the floor had done it before.”
The director or the client approves the structure from that page before anyone opens the timeline, and the editor builds the first assembly from it.
When the cut is done, read its transcript next to the approved script. Mandatory lines, legal wording, product claims and names should all be there as approved.
If your scripts keep on-screen text and spoken lines in separate columns, as in our two-column script template, the comparison goes quickly.
How this works in RecReview
In RecReview, a version with audio that goes through AI Review gets its transcript at the same time, with the language detected from the speech, and the transcript doesn’t use AI minutes beyond the review.
For a version uploaded without AI Review, anyone on the team can open the Transcript tab on the review page, pick Portuguese, English, Spanish or automatic detection, and generate it. That uses the workspace’s AI minutes for the length of the video.
The transcript comes in paragraphs, each with its timecode and, when the voices can be told apart, a label such as Speaker A. Click a word and the player jumps to it. While the video plays, the current word is highlighted and the list scrolls along.
Words the transcription was less sure of look slightly dimmer, the least certain ones in italics, and hovering over a word shows its time and confidence, so you know where to start the review pass.
The search box in the tab filters the transcript by word, and the app’s global search finds videos by what someone said in them.
If the task has a script and the client hasn’t seen the video yet, AI Review compares the speech with the script and points out lines that were dropped, added or swapped, each at its timecode.
The person who uploaded the version can switch the transcript to edit mode and fix any word with a double click. The subtitles are then rebuilt from the corrected text, unless someone already edited them by hand in the player, in which case those stay as they are.
The transcript exports as SRT or VTT with the same subtitles you see in the player, as TXT with a timecode at the start of each paragraph, or as JSON with the timing and confidence of every word.
