AI voice recorders and qualitative research
Dr Susanne Friese · · Research practice
AI Voice Recording: The future of how qualitative researchers will capture interview data and notes?
A new generation of very small AI-enabled recording devices has appeared over the past year. They attach magnetically to the back of a phone, clip onto clothing, or simply disappear into a pocket.
Press a button and they record. Once connected to the accompanying app, they can transcribe the recording, distinguish speakers, produce summaries and, in some cases, allow you to ask questions about the conversation. For qualitative researchers, this is potentially very useful.
Imagine finishing an interview and having a transcript available shortly afterwards, already separated into speaker turns. No SD cards, no separate recorder, no manual uploading to a transcription service. Some of these devices cost little more than a conventional digital recorder, and transcription plans are increasingly affordable.
I have been looking at four of them: Plaud, Pocket, Soundcore Work and EchoScrib.
Technically, I like the idea very much.
From a research-data perspective, however, there are some questions we need to ask before simply attaching one to our phone and starting the next interview.
What these devices can do The basic concept is similar across most of them. The recorder captures the audio locally. The associated app then produces a transcript and often also summaries, action points, mind maps or other AI-generated outputs.
For qualitative research, the most relevant feature after transcription is probably speaker separation. Anyone who has transcribed an interview knows how important this is. A transcript that does not reliably distinguish interviewer and participant is of limited use. Several of these devices now advertise speaker identification or speaker labelling.
Soundcore Work, for example, advertises speaker identification alongside transcription in more than 100 languages. Its privacy policy explains that the system analyses voice characteristics to distinguish speakers and can attribute text to speaker identities that the user has marked. (soundcore
soundcore.com/uk/products/d3200-soundcore-work-… )
Plaud has gone a step further. It allows users to create a profile for their own voice and to label other speakers. Its automatic speaker-labelling feature can then recognise frequent speakers in later recordings, and Plaud says recognition improves as more labelled recordings are processed. (Plaud
support.plaud.ai/hc/en-us/articles/54027338385177-… )
Pocket offers speaker labels and, on its Pro plan, advertises automatic detection of speaker names. (Pocket
I would nevertheless check speaker assignments carefully before using such a transcript analytically. Speaker diarisation — determining when one person stops and another begins — and identifying who that speaker is are not quite the same problem. Systems can still confuse speakers, particularly where voices are similar, people interrupt each other or speech overlaps. For research interviews, I would probably begin a recording explicitly: “I’m Susanne, the interviewer, and I’m talking today with Maria.”
That provides useful contextual information for the transcription system and can help with assigning names after speakers have been separated. But I would not rely on that alone. Where the software allows it, explicitly labelling speakers or creating a voice profile is preferable. This is an area where I expect considerable improvement over the next few years.
Four devices worth knowing about
Plaud Plaud is probably the most established of this new category of AI recorders. Recordings can remain on the device and associated app, with cloud processing invoked when transcription or summarisation is requested. Its current documentation is relatively detailed. Plaud says that when cloud synchronisation is disabled, recordings are uploaded for processing but are not retained on Plaud's servers afterwards. Its AI providers currently include OpenAI, Google and Microsoft under enterprise arrangements that Plaud says prohibit model training and require zero retention of processed content. Plaud also operates regional infrastructure. EU/EEA accounts are assigned to AWS Frankfurt, accounts in North and South America to AWS Oregon, and Asia-Pacific and Japanese users to regional infrastructure in Singapore and Japan. (Plaud
support.plaud.ai/hc/en-us/articles/49922008621337-… )
For qualitative research, that makes Plaud one of the devices to look into.
Soundcore Work Soundcore Work is particularly interesting from a data-minimisation perspective. By default, recordings and transcripts are stored locally and encrypted with AES-256. Audio is uploaded only when transcription is requested. According to Soundcore, when cloud storage is disabled, the uploaded audio and transcription are deleted after the result has been delivered. Files that have not successfully been retrieved can remain encrypted for up to seven days.
Soundcore also states that data are not used for model training without permission, and that transfers are encrypted. For research use, I find this architecture attractive because recording and permanent storage can remain local while cloud processing is temporary. (soundcore
soundcore.com/eu/products/d3200-soundcore-work-… )
Pocket Pocket is perhaps the most appealing from a pure usability perspective. It can record offline, synchronise later, generate transcripts, separate speakers and provide summaries and other AI functions. The free plan currently includes unlimited standard transcription, while the Pro version provides higher-accuracy transcription, automatic speaker-name detection and additional AI functionality.
Pocket also gives users some control over syncing. Its current trust documentation says recordings can remain on the device or be synchronised to its encrypted cloud. There is, however, an important regional issue: Pocket states explicitly that transcription and summarisation are processed on US-based cloud infrastructure.
For a researcher in the United States, that may not in itself be problematic. Institutional rules, informed consent, data sensitivity, encryption, retention and applicable state or sector-specific regulation still matter, but there is no international transfer merely because processing takes place in the US.
For researchers in the EU or EEA, another layer is added: transferring identifiable interview material to US infrastructure brings GDPR international-transfer requirements into the picture. That does not automatically make Pocket unusable. But I would want to examine the contractual and institutional arrangements before sending research interviews through its cloud workflow. (Pocket
EchoScrib EchoScrib currently does not seem to be suitable for interview recording. The device uses the AIREC app for AI-enabled transcription. The current Google Play disclosure for that app says that it may collect personal information, messages and audio, and — more concerningly — states that the data are not encrypted. (Google Play
play.google.com/store/apps/details ) That disclosure may not tell us everything about how every part of the system operates. App-store privacy information is supplied by the developer and can change.
But for qualitative interview data, the burden should not be on the researcher to guess what happens to the recording. Until the processing architecture, encryption and data-handling arrangements are explained more transparently, this would not be my choice for research involving confidential or identifiable material. (EchoScrib
The important issue is not the recorder The hardware is almost the least interesting part of this. If you search, you will probably find a few more companies who sell such a device. (Note: If you come across ABH Pocket, I would currently be cautious. Its public information provides little detail about where recordings are processed, which AI or transcription providers receive the data, how long data are retained, or whether they are used for model training.)
The crucial question is what happens after you press stop.
An interview recording is potentially highly sensitive research data. Even if the final transcript is anonymised, the original recording contains a person's voice and may contain names, workplaces, locations, third parties, health information, political opinions and numerous incidental details that will later disappear from the cleaned transcript.
So before using one of these devices for research, I would want answers to a number of fairly mundane questions: • Where is the recording stored? • Does it leave the device? • Where is it processed? • Who performs the transcription? • Which other AI providers receive the audio or transcript? • Is anything retained after processing? • Can cloud storage be switched off? • Is the material used for model training? • Is the data encrypted while being transmitted and while stored? • Can it genuinely be deleted? • And does my participant information and consent procedure cover this processing?
Those questions matter regardless of whether the manufacturer's website displays the words GDPR, HIPAA, SOC 2 or “enterprise-grade encryption.” These terms describe different things. SOC 2 or ISO 27001 concerns organisational and security controls. Encryption concerns technical protection. HIPAA applies to particular kinds of healthcare information and organisations in the United States. GDPR governs processing of personal data in a European legal context. None of these labels alone tells you whether a particular research project should use the service.
Location matters There is therefore no single answer to the question: “Can qualitative researchers use these devices?”
A researcher working in California, Amsterdam or London may use exactly the same recorder but face different legal and institutional requirements. For a US researcher whose data are processed in the United States, the primary questions may concern institutional approval, participant consent, data security, retention, state privacy law and — where relevant — sector-specific requirements such as HIPAA.
For researchers in the EU and EEA, GDPR requirements apply, including additional requirements where personal data are transferred outside the EEA. EU-based processing can therefore make the situation considerably simpler. UK researchers operate under UK GDPR and UK-specific international-transfer arrangements. And in all jurisdictions, a university, research organisation, client or ethics committee may impose stricter rules than the minimum required by law. Therefore you need to ask:
Is the complete recording and transcription workflow appropriate for my participants, my research project, my jurisdiction and my institution?
A very promising development — with some homework attached Despite these reservations, I think this new generation of recorders is worth watching. For qualitative researchers who conduct many interviews, there is something rather compelling about a device that sits on the back of your phone, records excellent audio, produces a speaker-separated transcript shortly afterwards and costs comparatively little.
The technology removes a substantial amount of administrative work between conducting an interview and beginning analysis. But precisely because that process has become so effortless, it is easy to overlook what has happened in between. The recording may have travelled through several technical systems before the transcript appears on our screen.
At the moment, Plaud and Soundcore Work are the two I would investigate most seriously for qualitative research, because they provide comparatively detailed information about storage, processing and retention and allow workflows that reduce persistent cloud storage. Pocket is technically attractive, particularly for US-based researchers, but its US processing architecture deserves closer examination for European research projects. EchoScrib/AIREC currently leaves too many questions unanswered for me to be comfortable using its cloud transcription with confidential interview data.
As with all AI generated output, also here speaker labels, transcripts and AI-generated summaries remain outputs to be checked — not research evidence to be accepted uncritically.