How to Transcribe an Interview: Audio to Text

To transcribe an interview well, work in order. Record it so that a machine can hear every speaker. Decide what kind of transcript you need. Let an automatic tool write a first draft, then check that draft against the audio before any quote leaves your hands. The draft saves most of the typing, and skipping the check is where the risk comes from. If the interview is for research, settle consent and data protection before you upload anything.

A good recording saves the most work: interview transcription starts before the interview

The recording decides how many errors the automatic draft will contain, so the cheapest hour of work is the one you spend on sound before the interview. Speech recognition gets worse as audio gets harder. OpenAI's Whisper model, tested as Whisper Large V2, had a word error rate of 2.7 percent on LibriSpeech clean, a set of read speech ¹. On CallHome, which records conversations between friends and family, the same model had 17.6 percent ². On the CHiME6 test set it had 25.5 percent ³.

Read speech is the easiest case. Real interviews are conversations, and conversations overlap, drift between rooms and hit background noise. The same tool can be accurate on one recording and unreliable on the next, so the practical goal is to make your recording look like the easy case. These steps help:

• Put a microphone close to each speaker. A laptop microphone across a table gives the tool more room noise and less voice. If your recorder can save one track per person, use that.
• Choose a quiet room with no fan, traffic or music, and close windows and doors.
• Run a short test before the real interview, with the same setup and a similar voice. Northwestern University's research guidance says it is "always a good idea to test the service you plan to use with a recording similar to your research data" to see whether its accuracy is good enough ⁴.
• Keep the original file uncompressed. The Library of Congress describes linear PCM as capturing audio without lossy compression, and says that its fidelity is enhanced by higher sample rates and sample sizes ⁵. Keep the recorder's native file rather than an exported MP3.

If the interview is by phone, read the CallHome figure as the closer comparison. If you already have a recording, our overview of how to transcribe audio to text covers the tools and steps, and the free audio to text tool for files works in the browser.

Choose the transcription style the use needs before you start

Decide the style of the transcript before you record, because it sets how long the work takes and what the transcript can support. A full verbatim transcript keeps every word, stutter, false start and pause. A clean verbatim transcript drops the fillers and false starts and keeps the words. The glossary entry on verbatim transcription and the entry on clean verbatim define the terms, and our guide to verbatim and clean verbatim transcription sets out the rules for each.

Research methods writers treat the choice as a matter of interpretation, not housekeeping. Oliver, Serovich and Mason describe two main approaches. In naturalism, every utterance is transcribed in as much detail as possible. In denaturalism, "idiosyncratic elements of speech (e.g., stutters, pauses, nonverbals, involuntary vocalizations) are removed" ⁶. Their argument is that the transcript is not a neutral copy, and they state it plainly: "we suggest that transcription is a powerful act of representation" ⁷.

Match the style to the question. A thematic coder usually needs clean text with reliable quotes. A journalist who prints a quote has to choose between a faithful and a tidied version, and should say which one the piece uses. Whatever you choose, write it down in your methods or your notes, because a reader cannot judge a transcript without knowing how it was made.

Style also decides the cost. The time figures in the next section mostly describe full manual transcription, which is the slowest route.

AI drafts most interviews well, but every quote must be checked against the audio

An automatic draft gets you most of the way, and the check gets you the rest. A useful comparison of automatic tools for interviews is the 2024 study by Wollin-Giering and colleagues. They tested nine tools on one English and one German interview, and they conclude that "Whisper performs best overall" ⁸. They also recommend reviewing the text of any transcript: "For any type of transcription, we recommend reviewing the text to ensure accuracy" ⁹.

Start with the cost of doing it by hand. Wollin-Giering and colleagues cite Gläser and Laudel, who found that verbatim transcription took "an average of six hours to transcribe one hour of interview material" ¹⁰. Kvale and Brinkmann estimated five hours per hour of interview ¹¹. Those are estimates from earlier research, not timed tests of today's tools.

The newer evidence comes from a test run by F4x, which sells its own automatic transcription service. Its study found that students "needed on average 6.3 times the interview duration" to transcribe by hand ¹². Correcting an automatic draft took them 5.1 times the interview duration, and the study reports a saving of "70 minutes of correction time per hour of interview material" ¹³. Because a vendor ran the test, treat the numbers as one data point.

Tools differ widely on the same recording. On the German interview, Whisper reached "above 70%" on the study's word accuracy measure, while Dragon's was "below 5%" ¹⁴. That is one recording and a proxy measure, not standard word error rate, so use it to see the spread and then test your own tool on your own audio.

Automatic tools also fail in ways a human would not. Whisper's authors list complete hallucination among its known failure modes, in which the model outputs a transcript unrelated to the actual audio ¹⁵. Errors can also fall unevenly across speakers. A 2020 study of five commercial systems found that the average word error rate was 0.35 for Black speakers and 0.19 for white speakers ¹⁶. Those systems predate Whisper, but the finding is a reason to check every speaker with the same care, and to check most closely where the draft looks thinnest.

For each interview, check the draft in this order:

1. Play the audio at normal speed while you read the draft. Skimming misses the errors that matter most.
2. Check every quote you will use, word for word, against the audio. Do not trust a draft quote because it reads well.
3. Check names, numbers, dates and negatives. One missed "not" can reverse the meaning of a sentence.
4. Mark unclear words in square brackets, such as [inaudible] or [unclear: possibly "fund"], rather than guessing.
5. Save the checked version and note which tool and model made the draft, so you can explain the method later.

Our comparison of human and AI transcription covers when a person should do the work instead. For the models themselves, see our guide to Whisper, ChatGPT and Gemini transcription and the guide to Whisper's AI. A draft is also measured by its word error rate, which compares the draft to a checked transcript.

Research interviews carry consent and data-protection duties

For a research interview, the first checks are legal and ethical, not technical. Four things set the rules: recording law, data protection law, your institution's approved tools and the terms of the service you use. Check all of them before you upload the recording. This section is general information, not legal advice, and your ethics committee or legal team has the final say.

Recording law varies by place. Under the federal Wiretap Act in the United States, it is generally not unlawful to record a conversation when you are a party to it, or when one party has given prior consent to the interception ¹⁷. California is stricter: Penal Code 632 punishes anyone who intentionally records a confidential communication without the consent of all parties ¹⁸. An interview held in California, or with someone in California, should be treated as needing everyone's consent. Our guide to recording consent laws covers the state and country rules, and the glossary explains all-party consent.

Under data protection law, a recording of a person's voice is personal data. The GDPR defines biometric data as personal data from specific technical processing of physical, physiological or behavioural characteristics that allows or confirms the unique identification of a person ¹⁹. Microsoft makes the practical point for its own service: audio of people speaking and its transcripts may be considered personal data or sensitive data under various privacy regulations and laws ²⁰. An ordinary interview recording is not biometric data unless you process it to identify someone, but it is personal data, so you need a lawful basis ²¹, limited access and a deletion date. Our guide to whether a voice recording is personal data goes into the detail.

The tool you use also has rules. OpenAI says that abuse-monitoring logs for its API are retained for up to 30 days by default ²², and other providers publish their own periods, which can differ. Ask any provider how long it keeps audio and text, whether it uses your content to train its models, and where it stores the data. The guide to AI training on voice data explains what to look for.

Institutions usually keep approved lists, and the lists differ. Northwestern's guidance says its Zoom recording option should not be used with data subject to HIPAA, or with data use agreements that prohibit use of Zoom or AI services ²³. The University of Utah warns that recorded or transcribed content may be subject to public records requests, and lists this as a concern with unapproved tools ²⁴. A consumer tool's defaults can also cause trouble. A federal class action filed in August 2025 alleges that Otter.ai's Otter Notebook service "by default does not ask meeting attendees for permission to record" ²⁵. The complaint also says the service fails to alert participants that recordings are shared with Otter to improve its artificial intelligence systems ²⁵. Those are allegations, and the case may have moved on, but the lesson for researchers holds: check the defaults before you press record.

Name the transcription service in your consent form and say where the recording is stored. That is advice rather than a legal rule, but it lets participants decide with the facts in front of them. The guide to voice data retention explains how long providers keep what they receive. Our question on whether you can use AI transcription for research interviews answers the short version.

Speaker labels and timestamps make quotes findable again

Speaker labels turn a block of text into a record of who said what. Timestamps link each line back to the audio, so you can check a quote in seconds. Many automatic drafts arrive without labels. OpenAI's speech-to-text guide describes a separate model for identifying who speaks, and says it should be used only when you need to identify who speaks during different parts of a recording ²⁶. If your tool does not label speakers, add the labels yourself during the check, and use the same names or codes throughout.

Use a label that fits the study: the role ("Interviewer", "Participant"), a code ("P04") or a pseudonym. Put a timestamp at the start of each speaker turn, or at regular intervals in a long answer, so that anyone can find the moment in the recording. The glossary entry on timestamps explains the format, and our guide to timestamps and speaker labels covers the options. Speaker identification by machine is harder than it looks: the entry on speaker diarization explains what the tools are guessing.

Transcribe a podcast

A podcast is the same job with a public audience, and the speaker labels matter more because listeners will quote the episode. Apple Podcasts shows transcripts for episodes, with a list of supported languages: "Transcripts are available with iOS 17.4 or later for podcasts in English, Danish, Dutch, Finnish, French, German, Italian, Norwegian, Portuguese, Spanish, and Swedish" ²⁷. If you publish the show yourself, you can supply your own file, and Apple's support page says: "Providing a VTT file allows you to identify every speaker with each line" ²⁸.

Our question on how to transcribe a podcast sets out the steps, and the glossary entry on WebVTT explains the file format. Check the episode against the audio before you publish, because a transcript that misquotes a guest is a correction waiting to happen.

Prepare the file for coding

Qualitative analysis software needs a clean, editable file. NVivo's help for version 14 on Windows says that "PDF files are created by importing PDF files. You cannot edit PDFs in NVivo" ²⁹. Keep your working transcripts as Word or plain text files, and export a PDF only for reading. Our question on preparing transcripts for qualitative analysis covers the import steps in more detail.

Anonymise while you transcribe, not afterwards. The CESSDA Data Management Expert Guide says: "Plan anonymisation at the time of transcription or initial write-up" ³⁰. Replace names with pseudonyms, use the same pseudonym for the same person in every transcript, and keep the key that links pseudonyms to real names in a separate, secured place. Changing a name in a single file after coding has started is how a project ends up with two people called "Sarah".

Keep the recording: the audio, not the transcript, is the primary record

The recording is the source, and the transcript is a file derived from it. A transcript can contain errors, can be edited for anonymisation and can be replaced. The recording is the evidence you check against, so keep it until your project no longer needs it, and keep the original file, not only a compressed copy.

The UK Data Service's recommendations for a 2023 anonymisation project show the working practice. Its guidance says: "A copy of the original file is to be made; all changes will be made to the new file and the original file is to be kept until all anonymization of the whole project is complete" ³¹. Edit a copy, leave the original untouched, and destroy the originals only when the project's data plan says so. The project described covered 38 interview transcripts ³¹, so the same procedure scales to a single interview.

Some interviews may be better served by careful notes than by a recording. The choice is part of your method, and your ethics approval should describe it. Once you do record, decide the retention and deletion dates before the first interview, and write them into the consent form and the data plan. Our guide to voice data retention covers what providers keep, and the guide to transcription software helps you choose a tool that fits the rules above.

The whole job, from recording to storage, runs in this order:

Each step narrows the risk for the next. A clean recording gives the tool a fair chance. A style chosen in advance tells you what to keep. A checked draft is one you can quote. Labels and timestamps let a reader find the moment, and anonymisation with the original kept lets you answer a question years later, from the audio, not from memory of a draft.