AI transcription for qualitative researchers: fieldwork done, notes ready

You've just finished a 90-minute semi-structured interview with a research participant who disclosed more than you expected — personal history, professional risk, details they've never shared in a formal context before. You're in a community centre in a rural town with no mobile signal. You have two more interviews this afternoon. The cloud transcription service you usually use isn't loading.

This is not an edge case. It is a typical afternoon of qualitative fieldwork.

Qualitative researchers — PhD students, postdoctoral researchers, social scientists, anthropologists, ethnographers — work in locations that cloud tools don't reach, with participant data that cloud tools shouldn't hold. The research ethics framework that governs their work was not designed with AI transcription in mind, and the tools that exist for AI transcription were not designed with research ethics in mind.

On-device AI transcription resolves both problems simultaneously.

The qualitative transcription problem

Manual transcription of qualitative interviews takes 6–8 hours per hour of audio. This is a well-established benchmark in qualitative research methods — the British Library transcription guidance and multiple research methods textbooks cite the same range. For a PhD student conducting 30 interviews of 60–90 minutes each, this represents 180–360 hours of transcription work — a month or more of full-time effort that produces no academic output, only the precondition for analysis.

Professional transcription services charge £1–2 per minute of audio. A 90-minute interview costs £90–180 to transcribe professionally. For a 30-interview study, that is £2,700–5,400 — typically the majority of a UK PhD student's annual UKRI stipend supplement, and far beyond what most fieldwork budgets accommodate.

Cloud AI services have reduced this to minutes. Otter.ai, Rev, and similar tools can produce a 90-minute transcript in under 5 minutes at a fraction of the professional cost. The accuracy is sufficient for qualitative analysis when reviewed against the original recording.

The problems: connectivity and confidentiality.

Fieldwork happens where cloud tools don't reach

Qualitative research is conducted in the places where the research questions live. That often means:

• Rural communities with limited broadband and no mobile signal
• Community halls, village pubs, participants' homes in areas with poor connectivity
• Overseas fieldwork locations in countries with limited or expensive mobile data
• Hospital environments where personal mobile hotspots are blocked
• Prisons, secure settings, and institutional environments with no personal device internet access
• Schools and colleges with firewalled Wi-Fi that blocks consumer cloud services

Cloud AI transcription requires a stable internet connection to upload audio for processing. In the locations described above, it doesn't work. The researcher either accepts the limitation and transcribes manually, or uploads the audio when connectivity returns — hours or days later, from a different location, with whatever risks that delayed transmission implies.

On-device transcription produces a full transcript in real time regardless of connectivity. The 90-minute interview is transcribed by the time the conversation ends. The researcher can review the transcript while the participant's contributions are still fresh.

Research ethics and participant confidentiality

Research involving human participants in the UK requires approval from a Research Ethics Committee (REC). The ethics application includes a data management plan that describes how participant data will be stored, protected, and eventually destroyed. For most UK research institutions, this is governed by the institution's data protection policy under UK GDPR.

Participant interview audio is personal data. When participants disclose sensitive information — health history, personal trauma, professional risk, political views, criminal history — that audio is likely to constitute special category data under GDPR Article 9, requiring the highest protection standard.

RECs are increasingly asking about AI transcription in data management plans. The relevant question is not whether AI transcription is used, but where the audio goes when it is processed.

The cloud transcription disclosure problem. Standard participant information sheets and consent forms describe how data will be stored and who will have access to it. They do not typically describe AI cloud transcription services, because most researchers using cloud AI have added this to their workflow without updating their ethics documentation. A participant who consented to an interview for a university research project did not necessarily consent to their audio being processed by a US AI company under US law.

This is not a technicality. Research participants — particularly those sharing sensitive or personally risky information — have a reasonable expectation that their disclosure goes to the researcher and, through the institution, to ethical and legal review. An AI company's servers are not within that expected boundary.

The GDPR data transfer issue. For UK research, transferring participant audio to a US company's servers involves a cross-border data transfer. US companies handling EU/UK data are subject to UK GDPR's transfer mechanisms — Standard Contractual Clauses or adequacy decisions. Most AI transcription services provide SCCs in their enterprise terms; the free and standard tiers of most consumer AI transcription services do not explicitly address UK GDPR research data obligations.

The REC framing for on-device processing. On-device transcription resolves the ethics disclosure problem cleanly. The data management plan can accurately state: "Interview audio is transcribed using on-device AI software. At no stage is audio transmitted to any external server or third-party processor. Transcripts are stored on the researcher's encrypted device and backed up to the institution's secure research data repository." This is a description RECs can assess and approve without requiring disclosure of a third-party data processor's terms of service.

Kuulo in qualitative fieldwork

The interview session. Record the interview with Kuulo on an iPhone. Transcription runs on-device in real time — no internet required, no upload, no waiting. Speaker diarization attributes the interviewer's questions to one voice and the participant's responses to another. For semi-structured interviews with probing follow-ups, this separation significantly reduces the time spent on transcript preparation.

After the interview. The full transcript is on the device. Generate a thematic summary — key topics the participant discussed, notable quotes, emerging themes relative to the research question. This is available before the researcher leaves the field site.

Multiple participants. Each participant's interview is a separate recording on the device, separately tagged and summarized. Across a 30-participant study, the researcher has 30 structured, searchable transcripts with thematic summaries — the pre-analysis foundation that typically takes weeks to construct from manual transcription.

Export. Transcripts export to text formats compatible with qualitative data analysis software: NVivo, Atlas.ti, MAXQDA, Dedoose. The on-device transcript goes directly into the analysis environment without any additional cloud processing step.

Speaker diarization for interview research

Qualitative interview transcripts traditionally require speaker identification: differentiating the interviewer's questions from the participant's responses. In a well-conducted semi-structured interview, the interviewer's voice is distinct, consistent, and a relatively small fraction of the total audio. Manual transcript preparation still requires the transcriber — or the researcher reviewing a professional transcript — to add speaker labels.

Kuulo's on-device diarization does this automatically. The result is a transcript that already separates interviewer from participant, reducing the review time significantly and producing a cleaner starting point for analysis.

For focus group research — where multiple participant voices need attribution — diarization becomes more complex. Kuulo handles a moderate number of simultaneous speakers; for a focus group of 8 participants, speaker attribution is less reliable than for a one-on-one interview. This is an honest limitation. For focus group data, manual speaker attribution remains necessary; Kuulo handles the transcription foundation and the researcher adds speaker labels during review.

The cost case

Professional transcription at £1.50/minute × 90 minutes = £135 per interview. Across 30 interviews: £4,050.

Kuulo: free to start.

The saving is not marginal. For a UK PhD student on the UKRI stipend (approximately £19,237 in 2026), £4,050 represents more than 20% of annual income. For a postdoctoral researcher on a fixed-term PDRA contract with a limited fieldwork budget, professional transcription is often the cost that forces compromises on sample size.

Removing the transcription cost bottleneck changes what is feasible in qualitative research design. A researcher who can afford to conduct 40 interviews rather than 20 — because AI transcription costs nothing rather than £5,400 — has access to a richer dataset and a more convincing study.

A research ethics application paragraph

For researchers who need to describe on-device AI transcription in an ethics application or data management plan, the following is accurate and can be adapted:

"Interview recordings will be transcribed using Kuulo, an on-device AI transcription application installed on the researcher's password-protected iPhone. All transcription processing occurs locally on the device; at no stage is audio transmitted to external servers or processed by third-party cloud services. Transcripts are stored on the device's encrypted local storage and backed up to [institution's secure research storage platform] within [timeframe]. The original audio recordings will be deleted upon transcript verification in accordance with the data minimisation principle. This approach ensures that participant audio remains within the researcher's custody and is not disclosed to third parties."

This description is accurate. It is consistent with UK GDPR data minimisation and security requirements. It does not require the researcher to explain or justify any third-party AI company's data handling practices, because there is no third party.

The field researcher's practical checklist

Before fieldwork begins:

• Install Kuulo on an encrypted, password-protected iPhone
• Update ethics documentation to describe on-device transcription
• Test transcription quality with a sample recording in a similar acoustic environment
• Ensure sufficient storage for the number of interviews planned (90 minutes of audio ≈ 1GB; 30 interviews ≈ 30GB; verify device storage)

During fieldwork:

• Record each interview separately, tagged by participant code (not name)
• Verify transcript at the end of each day while audio is still fresh
• Export verified transcripts to secure institutional storage

After fieldwork:

• Delete original audio from device once transcripts are verified
• Archive transcripts per the institution's data management policy

The qualitative research tradition is built on trust between researcher and participant. That trust extends to how the participant's words are handled after the interview ends. On-device processing is the architecture that honours that trust — not because it is required by a compliance framework, but because it is what the trust relationship actually means in practice.