Home → Blog → How to transcribe an interview without retyping it by hand

How to transcribe an interview without retyping it by hand

An hour of interview takes four to six hours to type by hand. Automatic transcription cuts that to minutes of machine time plus twenty minutes of checking — but only if the recording and the checking are done right, and that is where most of the quality is decided.

Most of the accuracy is decided before you press record

No model recovers information that was never captured. Three things matter more than any parameter you will set later:

• Put the microphone near the mouths. A phone lying between two people on a café table records mostly the café. The same phone held at chest height records the interview.
• Kill background noise you control. Air conditioning, an open window, background music — each of them costs you accuracy on every sentence.
• Record two tracks if you can. Two phones, or a call recorded per-channel, gives you the speaker split for free later. This is the single highest-value decision in the whole workflow.

Recording on a laptop through a cheap external microphone beats a phone in a pocket by a wide margin, and costs nothing to arrange.

The transcription itself

One request. Ask for segments with timings — for an interview they are what makes quotes checkable:

from openai import OpenAI

client = OpenAI(base_url="https://voicesscribe.com/v1", api_key="your key")

with open("interview.mp3", "rb") as f:
    r = client.audio.transcriptions.create(
        model="whisper-1", file=f,
        response_format="verbose_json",
        language="en",
        prompt="Interview with Dr. Elena Marchetti, epidemiologist, on wastewater surveillance.",
    )

for s in r.segments:
    m, sec = divmod(int(s.start), 60)
    print(f"[{m:02d}:{sec:02d}] {s.text.strip()}")

The prompt is where the interviewee's name, their field and the specialist terms go. Interviews are dense with proper nouns, and proper nouns are exactly where recognition slips — a one-line hint fixes most of them.

Two voices in one file

A single mixed recording does not come back labelled by speaker. That is a genuine limitation, and it has practical workarounds:

• Separate tracks. Transcribe each track on its own and you get a perfect split, with timings you can interleave.
• Separate channels. If the recorder put each person on a channel, split the file first: ffmpeg -i in.wav -map_channel 0.0.0 a.wav -map_channel 0.0.1 b.wav
• Mark it up afterwards. In a two-person interview the turns alternate predictably, and the interviewer asks the questions. Reading through once and adding names is a few minutes for an hour of text.

For most interviews the second and third options are enough. Do not build a diarisation pipeline before you know you need one.

Checking the result

Automatic transcription is accurate on ordinary speech and least reliable exactly where it matters most for an interview: names, numbers, organisations and specialist terms. A twenty-minute pass focused on those four categories catches nearly everything.

The timings make this practical — for any sentence you doubt, jump to that second of the recording and listen. This is also how you avoid the worst failure mode in published interviews: a quote that reads well and was never said.

# find the places worth listening to yourself
for s in r.segments:
    if any(w[0].isupper() for w in s.text.split()[1:]) or any(c.isdigit() for c in s.text):
        print(f"{int(s.start)//60:02d}:{int(s.start)%60:02d}  {s.text.strip()}")

From transcript to what you actually needed

The transcript is raw material. What gets used is usually one of three things:

• Quotes with timestamps — for an article, checkable against the recording.
• Themes across many interviews — for research, where the same questions were asked ten times.
• A summary for people who were not there — for a team that needs the conclusions, not the transcript.

All three can be drafted by a language model from the text, with one rule that keeps them honest: ask explicitly for quotes verbatim and for nothing that is not in the transcript. Then check the quotes against the timings.

Consent and storage

Two things worth settling before the recorder starts, not after:

• Tell the person they are being recorded and that the audio will be machine-transcribed. In many jurisdictions the first part is a legal requirement; the second is simply fair.
• Know how long the audio lives. Here, recordings and transcripts are deleted automatically after 24 hours — long enough to download the result, short enough that there is no archive to leak.

For research work with an ethics review, that retention window is usually the detail the committee asks about.

Frequently asked questions

Related reading

• How to transcribe meeting recordings and get usable notes — Turn Zoom, Teams and Google Meet recordings into searchable text and structured notes: where the file lives, how to transcribe it and how to extract decisions.
• Who said what: separating speakers in a transcript — Why a single mixed recording does not come back labelled by speaker, and four practical ways to get the split anyway — from multitrack recording to channel separation.
• How to improve speech-to-text accuracy: eight practical fixes — Eight changes that actually move transcription quality: audio preparation, language hints, the prompt parameter, chunking, formats, and how to measure word error rate.