Transcribe an interview without uploading it
Drop the recording in and get a timestamped, speaker-labelled transcript produced entirely on your machine. Nothing is uploaded, so a promise of confidentiality made to a participant stays a promise you have actually kept.
Built for the way interviews are really worked with: click a line to hear it, fix what was misheard, label who is speaking, then export to text for writing up or CSV for coding.
Language and translation need the Multilingual model.
Drop an audio or video file
or , paste with Ctrl + V, or
- MP3
- MP4
- WAV
- M4A
- MOV
- WEBM
- OGG
- FLAC
- AAC
The file is read by this page and never uploaded — no account, no size limit but your own RAM.
Preset: balanced English model, plain-text output. Add speaker labels once the transcript appears.
Download
Click a line to jump the audio there. Click the text to fix a word — your edits go into every download.
How it works
- Open this page — the tool is already set up for “Research interviews”.
- Drop the file in, browse for it, or paste it with Ctrl+V.
- The model runs on your own device; the transcript appears as it goes.
- Correct anything misheard, then download text, subtitles or notes.
The consent problem with cloud transcription
Most interview consent forms say something close to "the recording will be heard only by the research team" or "will not be shared with third parties". Uploading that recording to a transcription service is a third-party disclosure, and it has happened the moment the upload completes — before anyone has read a privacy policy or thought about retention. For university research this is usually the exact thing an ethics committee asked about; for journalism, it can put a source in a database the journalist does not control.
The technical position here is different in a way that is easy to verify: there is no upload endpoint on this site. The model is downloaded to your browser and the audio is processed by your own processor. You can open the network tab and watch, or disconnect from the internet after the model is cached and transcribe with the network off entirely.
Speaker labels done honestly
Whisper does not identify speakers. Tools that claim automatic "Speaker 1 / Speaker 2" labels are running a separate diarization model, and on two people with similar voices or any crosstalk it produces confident, wrong attributions — which in an interview transcript is worse than no labels at all, because a misattributed quote can survive into a publication.
So the labelling here is manual and fast: click a line, type the name, and it applies to that line and every one after it until the next label. A turn-taking interview usually needs a handful of clicks. The labels appear in the text, Markdown, CSV, JSON and subtitle exports.
Coding and analysis
The CSV export gives one row per segment with start and end times, the speaker and the text, escaped properly so quotes and commas survive Excel. That imports cleanly into a spreadsheet for open coding, or into NVivo, MAXQDA, Dedoose and Taguette via their delimited-text importers.
For a first pass, the JSON export carries the same structure with exact times, which is easier to script against if you are building a corpus. Both use your corrections, not the raw model output — so the version you code is the version you checked.
Things that save a re-run
- Label speakers before exporting; the labels are written into every format.
- Fix a participant’s name once, then use the search box to find and correct the rest.
- For a two-hour interview, transcribe in one pass and use the timestamps to navigate rather than scrubbing the audio.
- Record with the microphone closer to the interviewee than to you — their words are the ones you need verbatim.
Read the transcript against the audio before you rely on it. Every speech model — this one and the paid cloud ones — mishears names, numbers and crosstalk, and it does so confidently. The player and editor here exist so that check takes minutes rather than an afternoon.
Frequently asked questions
Does this satisfy my ethics approval or data-management plan?
That is for your committee or supervisor to say, and it depends on what you undertook. What can be stated technically is that the recording is not transmitted anywhere: it is read by the page and processed by your own device, so no new party gains access to it. That is usually the distinction such plans turn on, but confirm it against your own approval rather than this page.
Can it separate the interviewer from the participant automatically?
No, and it does not pretend to. Whisper has no speaker diarization. You label speakers manually — one click per turn change — which is both more accurate and more honest than an automatic guess that will misattribute quotes during crosstalk.
How accurate will it be on a real interview recording?
On a quiet room with a decent microphone, good enough that you are mostly correcting names and specialist terms. On a phone in the middle of a table with two people interrupting each other, noticeably rougher. Plan to read it against the audio either way — that is true of every automatic transcription tool.
Can I export it for NVivo or MAXQDA?
Use the CSV export: one row per segment with start time, end time, speaker and text. That imports into every major qualitative analysis package, and into a spreadsheet if you prefer to code there.