Learn · Qualitative research

Transcribe research interviews locally with Whisper

Whisper can transcribe research interviews on your own computer. After installation it works offline, so recordings stay on the machine. This guide covers setup, model choice, transcription, correction, and the change log.

By Heidi Andersén, MD, PhD. Updated 16 July 2026.7 minute read

What to expect with real interview audio

So calibrate interview transcription by language, not by hype:

A machine transcript you correct on your own computer is a normal research task. A polished-looking transcript from a cloud service, with the audio already elsewhere, is a methods problem wearing a convenience costume.

Why local transcription is easier to defend

Interview recordings are directly identifying data. Voice is identity. Keeping transcription on your own computer means no participant audio is uploaded to a third-party service, which is usually far easier to justify during ethics review. Offline transcription makes the disclosure question disappear instead of answering it.

What you need

• A Mac, Windows, or Linux computer with 8 GB of memory. No account, no license, no internet after setup.
• About 3 GB of disk space for the best model (smaller options below).
• Your recording as a file: mp3, m4a, wav, and most other formats work.

Step 1 — install Whisper · about 10 minutes

Install Python 3 from python.org. On Windows, tick "Add python.exe to PATH" in the installer.

Then open the command window. On a Mac: press cmd + space, type Terminal, press Enter. On Windows: open PowerShell. Paste this line and press Enter:

pip3 install faster-whisper

That is the whole installation. The speech model downloads itself the first time you transcribe, then lives in a cache on your machine and works offline.

Step 2 — pick a model size

Step 3 — run the transcription

Copy the script below into a plain-text file called transcribe.py, saved in the same folder as your audio. Then, in Terminal, move into that folder (type cd with a space, drag the folder into the window, press Enter) and run:

python3 transcribe.py

On Windows, type python transcribe.py instead — the python.org installer does not create a python3 command.

from faster_whisper import WhisperModel

model = WhisperModel("large-v3", device="cpu", compute_type="int8")

segments, info = model.transcribe(
    "interview1.mp3",
    language="sv",        # force the language; leave out to auto-detect
    initial_prompt=(      # your study's words: places, names, clinical terms
        "Vasa, Korsholm, cellgiftsbehandling, pemetrexed, kontrollbesök"
    ),
)

with open("interview1_raw.txt", "w", encoding="utf-8") as out:
    for s in segments:
        t = int(s.start)
        stamp = f"{t//3600:02d}:{t%3600//60:02d}:{t%60:02d}"
        out.write(f"[{stamp}] ?: {s.text.strip()}\n")

print("done - now correct it against the audio")

Two lines in there earn their keep:

• language= forces the language. Do this for bilingual interviews; auto-detection can flip mid-file when speakers code-switch.
• initial_prompt= takes a short line of your study's own vocabulary. The model spells proper nouns and domain terms far better when it has seen them once. Municipalities, drug names, local institutions: put them here.

Transcription runs at roughly the speed of the recording on a normal laptop. A one-hour interview takes about an hour. Start it, work on something else.

Step 4 — correct the transcript and log every change

The ? on every line is deliberate. The model does not know who is speaking, and it should not pretend to. The human pass is part of the method:

1. Listen with the transcript beside you. Fix misheard words. Names, terms, numbers, and doses get checked every time.
2. Replace every ? with the speaker (Interviewer, P01). Pseudonymize names in the text as you go.
3. Keep a change log, one row per correction. Minutes of work, and it turns "transcripts were checked" from a claim into a record:

Do not tidy how people actually speak. Hesitations, repairs, and dialect are data. The corrected transcript, with its change log, is the citable object.

What offline transcription does not do

• It does not separate speakers automatically. You assign them, which for interview data is minutes of work. Cloud services do offer speaker separation — the trade is that your audio leaves your machine.
• It does not make the transcript true. It makes a draft; the human pass makes it usable.
• It does not replace your ethics board or your judgment. It removes one disclosure they would otherwise have to weigh.

Related guides

More local-first qualitative research guides are on the way: anonymizing transcripts, Obsidian for qualitative analysis, AI-assisted coding with human review, and building an audit trail. They will appear on the Learn page.