Interviews are the case where transcription earns its keep — an hour of audio is several hours of typing — and also the case where the usual advice breaks down. Two people talking, often over each other, frequently under some obligation not to hand the recording to a third party.
Here is what works, and where you will hit a wall.
The wall: speaker labels
Whisper produces one stream of text. It does not tag who is speaking. Two people in conversation come out as continuous prose, with a change of speaker looking exactly like a change of sentence.
That is not a limitation of any particular tool. Separating speakers is a different task, called diarization, handled by a separate model. Services that show “Speaker 1 / Speaker 2” are running two models and stitching the results.
What to do about it:
- Interviewer questions are usually recoverable. You know what you asked. Reading the transcript with your question list beside it, the boundaries are mostly obvious.
- Mark turns as you go. If you can influence the recording, a half-second pause before each answer gives you a punctuation cue to find later.
- Two mics, two files. If the setup allows separate tracks, transcribe each file separately. You get perfect speaker separation for free, because each file only contains one person.
That last one is the real answer for anything where attribution matters.
The other wall: crosstalk
When two people talk simultaneously, the model does not produce two overlapping lines. It produces one line that blends both, and it reads as a grammatical sentence neither person said.
There is no fix at the transcription stage. If your recording has significant crosstalk, plan to check those passages against the audio — they are usually the most quotable moments, which is precisely why it matters.
What browser transcription is good at here
The confidentiality piece is where local processing changes the calculation, not just the price.
A recording of a source, a patient, a client or a research participant often carries a commitment about who can access it. “We delete files after 24 hours” and “the file was never sent anywhere” are different promises — the first is a policy, the second is an architectural fact. If your consent form or NDA says the recording will not be shared with third parties, uploading it to a transcription service is a disclosure, regardless of that service’s retention policy.
Running the model in the browser means the audio is read from disk, decoded, and processed in the same tab. Nothing is transmitted. You can verify this yourself: open your browser’s network panel, transcribe a file, and watch that no request carries your audio.
What an hour of interview actually costs you in time
Realistic expectations, because this is where browser tools get oversold.
The model downloads once — 73 MB for English, 237 MB for most other languages — and is then cached. In one measured run, a 13-second clip took 25.8 seconds end to end, of which 12.2 seconds was the one-time model download. That fixed cost disappears on every subsequent file.
After that, transcription runs at roughly a quarter to a half of real time on a typical laptop CPU, so a one-hour interview lands somewhere in the 15–30 minute range. You can leave the tab open and do something else; it does not need attention.
If that is too slow and the material is not confidential, a server-side service will be faster. That is the honest trade: local costs you time, uploading costs you the confidentiality claim. For most interview work the first is cheaper.
A workflow that holds up
- Record at 16 kHz mono or better, mic close to whoever is speaking. Audio quality beats every other variable including model size.
- Separate tracks if you can. It solves speaker labels completely.
- Transcribe, then read once against the audio at 1.5×. You are checking names, numbers, and anything you intend to quote.
- Fix proper nouns first. They are where errors concentrate — the model substitutes a plausible word rather than leaving a gap, so mistakes read as confident.
- Keep the original recording. For anything that could be disputed, the transcript is a working document, not the record.
Ready when you are — transcribe an interview with the file staying on your machine.
Common questions
How accurate is it on a two-person interview? On clean audio with little crosstalk, close enough that editing beats typing by a wide margin. Accuracy degrades with overlap, distance from the mic, and strong accents rather than with interview length.
Can it handle an hour-long file in one go? Yes, subject to your device’s memory. Long files are processed in overlapping 30-second windows rather than loaded whole. On a phone, expect to hit limits sooner than on a laptop.
What about interviews in other languages? Supported, but set the language explicitly, and check which languages are reliable first — accuracy varies more between languages than most tools admit.
Is this good enough for a legally admissible transcript? No. Certified transcripts require a human transcriber attesting to accuracy. Use this to draft, not to certify.