Transcribing interviews is one of the slowest parts of qualitative research. An hour of conversation can take four to six hours to type up by hand, so it is no surprise that researchers turn to automatic transcription.
The catch is where the audio goes. Most transcription services upload it to their servers. Interview recordings are among the most sensitive data a project holds: participants’ voices, their stories, often their names, workplaces and health. Sending them to an outside company can conflict with your consent forms, your ethics approval or your data management plan.
This guide shows how to transcribe interviews on an Android phone, with the audio never leaving it.
Check what your approval allows
Before choosing any tool, look at what you have already promised:
- Your consent form. Did you tell participants who would hear or process the recording? “Transcribed by the research team” is not the same as “uploaded to a commercial transcription service.”
- Your ethics or IRB approval. Many approvals name how recordings will be stored and transcribed, and require a new approval to change it.
- Your data management plan, and your institution’s rules on storing personal data, especially if participants are in the EU or UK, where the GDPR applies.
On-device transcription is usually the easiest option to fit within these, because no third party receives the data. Still check with your ethics committee or data protection office if you are unsure; they are the authority, not a blog post.
Step 1: Record the interview, or import it
With Offline AI Meeting, put the phone between you and the participant and tap record. It records for as long as the interview runs, with the screen off, and plays a tone at the start and end, so the participant knows exactly when recording is on. A notification shows the whole time.
If you recorded on a dedicated recorder or over a video call, move the file to your phone and import it: Import → Choose a file. MP3, M4A, WAV, Opus, FLAC and MP4 all work.
Get consent on the recording. Start recording, then confirm the participant agrees to be recorded. It is a useful backup to the signed form.
Step 2: Transcribe on the device
Transcription starts automatically and runs on the phone’s processor. Once the models are downloaded, it works in airplane mode, which is an easy way to show yourself, or a committee, that nothing is transmitted.
Choose the model in Settings → Transcription model. High is slower than Base but much better with accents, quiet speakers and background noise, and it is free. The models understand nearly 100 languages, so interviews that aren’t in English work the same way.
Step 3: Label interviewer and participant
With Pro, each line is labelled by speaker. Rename “Speaker 1” and “Speaker 2” to your own codes, such as “Interviewer” and “P07”, and every line updates. If the app hears the wrong number of voices, use the Fewer speakers and More speakers buttons to correct it.
Where both people talk at once, the label is a guess; the app marks this rather than hiding it. Long-press a line to reassign it, either for that line alone or for the whole turn.
Step 4: Check it against the audio
Automatic transcripts are a first draft. For research, read every transcript against the recording:
- Tap any line to play the audio from that point.
- Long-press a line → Edit text to fix a misheard word, name or number.
A note on verbatim transcripts: the speech model tends to write clean, readable sentences. It often leaves out fillers like “um” and “uh”, false starts and repetitions. If your method needs strict verbatim transcription, for example conversation or discourse analysis, add those back where they matter while you check.
Step 5: Anonymise before anything leaves the phone
Export is where data leaves your control, so de-identify first. With Pro, go to Settings → Redact before export and choose what to hide:
- Speaker names, exported as “Speaker 1”, “Speaker 2”
- Email addresses
- Phone, card and ID numbers
- Custom words: participant names, workplaces, towns, anything that could identify someone
You can also long-press any single line and choose Hide from exports, for a passage the participant asked to withdraw. Exports show [redacted] in place of hidden content; the original transcript on your phone is not changed.
Automatic redaction catches patterns, not context. A participant describing “the only paediatric nurse in our village” is identifiable without a single name in the sentence. Read the anonymised export before you share it.
Step 6: Export for analysis
Export in the shape your analysis needs:
- Word (.docx), which qualitative analysis tools such as NVivo and ATLAS.ti can import, as can most coding workflows
- CSV, one row per line with start time, end time, speaker and text, for spreadsheets or your own scripts
- Plain text (free), PDF, Markdown, or SRT subtitles if you also have video
Exports go only where you send them: saved to a folder you pick, or shared to an app you choose.
Keeping the recordings secure on the phone
- Use a screen lock. Android encrypts its storage, but the protection depends on a PIN, password or pattern.
- Recordings are kept in the app’s own storage and are left out of cloud backups.
- Delete when your plan says so. Many data management plans set a date to destroy original audio after transcription. Deleting a recording in the app removes it from the phone.
- Move transcripts to approved storage. If your institution requires research data on a managed drive, export the anonymised transcript there and delete the phone copy.
Summary
On-device transcription gives you most of the time savings of a cloud service, without handing participant audio to anyone. Record or import, transcribe offline, label and correct, anonymise, then export for analysis.
Get Offline AI Meeting on Google Play and try it on a test recording with airplane mode on.
This article is general guidance, not advice on your specific ethics approval or data protection obligations. Follow your institution’s requirements.