Konthora

Knowledge Center

How to Transcribe an Interview

An interview transcript has a different job from a podcast transcript. You are not publishing the whole thing — you are mining it for a handful of accurate quotes. That changes what you need: tighter timing, clean paragraph breaks, and a transcript you can trust enough to publish without a fact-check pass.

Accuracy you can actually publish

Automatic transcription is good enough to find a quote and accurate enough to check quickly. It is not accurate enough to skip checking. The failures are predictable: brand names, surnames, job titles, numbers, and anything said while someone laughs or talks over the speaker.

Read the transcript against the audio once, marking anything you plan to quote. If a quote will appear in a published piece, verify it word for word. This is faster than it sounds, because you are reading two short things side by side rather than re-listening to forty minutes.

If a quotation is legally sensitive — anything in a recorded interview that might be contested — transcribe it, check it, and keep the audio. The transcript is a finding aid, not evidence.

Two speakers on one track

Konthora does not label speakers. For an interview that means you are attributing lines yourself, and the fastest way to do that is to transcribe with word-level timestamps so you can jump to a moment in the recording without scrubbing blind.

If you recorded the interview on two machines, you have a genuine advantage: you already know which half of the audio belongs to which person, so attribution is a matter of cutting rather than inferring. Merge them in order and transcribe the combined file.

A single combined track is harder. Look for the practical tells — pitch, pace, and turn-taking rhythm are usually enough to follow a two-person conversation, but budget time for it.

Finding the quote worth using

Sentence timestamps make the transcript searchable in a way that matters here: you can skim the text for a phrase, note the timestamp, and jump straight to that point in the recording to check the delivery. A quote that reads well in text may sound wrong in the actual sentence, and the timestamp is how you find that out quickly.

Paragraph mode is better when you are looking for a passage rather than a single line, because it keeps the surrounding sentences together and the context usually determines whether a quote stands up.

Word mode is the right choice only when you need subtitles or karaoke-style alignment for a published clip.

Advice that happens before transcription

The quality ceiling is set at recording time, not at transcription time. A single mic in the middle of a table beats two mics at opposite ends of a room, and a lav on the speaker beats a laptop microphone by a wide margin.

Ask your subject to spell names and terms you will need. A five-second confirmation at the start of the recording removes the most common transcription error before it happens.

If the recording is already made and rough, do not discard it. Trim the worst noise, run the transcription, and correct the transcript by hand. That is usually faster than re-recording and it is always better than publishing an inaccurate quote.

Step by step

1

Consolidate the recording to a single file

Interviewers often record the local end and the remote end separately. Combine or choose the better track first — the two usually have different levels, and the weaker one is where errors concentrate.

2

Trim the pre-roll and dead air

Long silences and tape-room noise at the head and tail of the file do not produce useful text and only slow the job down. Cut them before upload.

3

Upload the recording in sentence mode

Sentence mode is the right default for interviews. It gives you one timestamp per sentence, which is exactly the granularity you need to find a quote later and verify it against the recording.

4

Transcribe, then read it against the audio once

The transcript is a first pass, not a legal record. Read it against the audio and correct anything that matters for your piece. Proper nouns and numbers deserve the closest attention.

5

Export TXT to read, JSON if you are scripting

TXT for reading and copying quotes. JSON carries per-word timing, which is useful if you are building a searchable quote index or aligning audio to text in a player. Export as soon as the result is ready: results are held for 60 minutes, and a verified interview transcript that expires before you have written up your quotes is an afternoon you will not get back.

Open the transcription tool

Frequently Asked Questions

Does Konthora label who is speaking in an interview?
No. There is no speaker diarisation, so lines are not attributed. Use word-level timestamps to follow a line back to the recording, or merge two separately recorded tracks in order so you already know which half belongs to whom.
How accurate is the transcript, and should I fact-check it?
It is accurate enough to locate a quote and then verify it quickly, but you should read it against the audio once before publishing. Names, job titles, numbers and anything said over background noise are the usual failure points.
Can I transcribe an interview longer than 10 minutes?
Media is capped at 100 MB and 10 minutes per file, so longer interviews need splitting. Split at topic changes and overlap the joins by about a second, since models can drop the first words of a file.
What if I recorded the interview on two devices?
Combine the two tracks in order, or transcribe them separately. Because you know which half of the audio belongs to each person, attribution becomes a matter of cutting rather than guessing, which is the most reliable way to work.