UX and academic researchers
Interview transcription software that keeps the recording and the transcription on your own Mac
Ghosty Notes transcribes participant interviews with a model running on your own Mac, with no bot in the call and no upload of the recording.
Ghosty Notes runs on a Mac with Apple Silicon. Intel Macs cannot run the on-device transcription model and are not supported.
no notetaker in this list
what the other side sees while you record
A research interview is a promise. The participant agreed to talk to you, at your institution, for the study you described to them in the consent form. Upload that session to a transcription service and a company they have never heard of receives an hour of their voice discussing whatever your study is about, which is not the arrangement they agreed to, whether or not anyone on the project treats it as a change.
Ghosty Notes removes that step. It records through the Mac's own system audio, so no attendee appears in Zoom, Teams or Meet, and speech becomes text using a model that runs on your machine. The audio never leaves the device on any plan. Transcripts are written as markdown files into a folder you choose, which suits what happens next: markdown imports cleanly into MAXQDA, ATLAS.ti, NVivo or Delve, so the coding stays in whichever tool your team already knows.
On-device transcription, unlimited interviews and a local search server are free, with no minute cap. Pro is $5.00 a month or $54.00 a year and adds calendar sync with Google and Outlook, hosted AI models, faster chat across many interviews, Obsidian sync and speaker identification.
What your ethics review is actually asking about transcription
Institutions have started naming these tools. A Loughborough University data-privacy notice dated 6 February 2026 warns that tools "such as otter.ai and MeetGeek introduce significant risks, especially when personal or sensitive data is involved", and directs staff to "primarily use the transcription features already available in both Microsoft Word and Microsoft Teams". That is one university and not a sector-wide rule, but it is worth reading for what it reveals. The objection is not to machine transcription, since the sanctioned alternative is also machine transcription. It is to sending the recording to a third party the institution has not assessed and holds no contract with.
Litigation runs along the same line. In re Otter.AI Privacy Litigation, case 5:25-cv-06911, was filed in the Northern District of California on 15 August 2025. It is pending and nothing about it is decided, so treat it as a live case rather than a conclusion about anyone's conduct. Its existence still matters to you, because a participant recorded on your project is exactly the kind of person on whose behalf that sort of claim gets brought, and because your institution's legal office reads the same filings.
The practical difference is where a vendor risk assessment comes from in the first place. Your institution reviews a transcription supplier because that supplier receives your data: it holds it, it has staff, a retention policy and a jurisdiction, and someone has to assess all of that before the data moves. Software that processes the recording on the researcher's own laptop puts no recipient into that chain, so it does not enter the review in the same shape. Be careful with that sentence in both directions, because it exempts you from nothing. Your ethics committee or IRB still governs the project, many institutions require approval of any software touching participant data wherever it runs, and your data management plan still has to describe what you do. What changes is the answer to "who else gets a copy", not whether you have to answer it.
That answer is now part of the paperwork. Ethics applications increasingly ask you to describe AI transcription as a data-sharing practice, and consent forms are being rewritten to name the tool, because a participant who agreed to speak with a researcher did not thereby agree to a commercial processor holding the recording. A workflow where the file stays on a machine you control is a short paragraph to write in an application and a short thing to say out loud to a participant who asks what happens to the recording afterwards. Neither of those is a substitute for asking, and both are easier when the answer is short.
- Loughborough University data-privacy notice, 6 February 2026
- Names otter.ai and MeetGeek as introducing significant risks where personal or sensitive data is involved, and points staff to the transcription already in Microsoft Word and Teams.
- In re Otter.AI Privacy Litigation, 5:25-cv-06911 (N.D. Cal.)
- Filed 15 August 2025. A pending case, with nothing decided.
Across a round of fieldwork
Say what happens to the recording
In the consent conversation the whole chain is one sentence: the audio is recorded and transcribed on this laptop and is never uploaded. Participants who ask about AI transcription want exactly that, and it is the same sentence your ethics application needs.
Run the session with nothing in the room
No bot appears in the participant list, so nobody asks who the extra attendee is and the first five minutes go on your questions instead of on explaining a tool that is not part of the study.
Take the transcript into your analysis tool
Each interview lands as a timestamped markdown file in the folder you chose. That imports into MAXQDA, ATLAS.ti, NVivo or Delve as plain text, because coding, memoing and intercoder work stay in the tool you already use. Ghosty Notes does not code.
Ask questions across the whole round
The MCP server lets Claude, ChatGPT or Cursor search your transcripts locally. "Which participants mentioned the export button" points you into the data before you have coded it, and it does not code it for you or claim to.
What researchers push back on
- Does using this get me out of ethics review?
- No, and anyone telling you otherwise is wrong. Your IRB or ethics committee governs the project wherever the processing happens, many institutions require approval of any software that touches participant data, and your data management plan still has to describe your workflow. What changes is one answer inside the review: there is no external recipient of the recording for anyone to assess.
- Our institution tells us to use Word and Teams for transcription.
- Then follow that, or get this approved before it touches participant data. An institutional instruction is a real constraint and a landing page does not override it. The argument to put to your data protection officer is architectural rather than promotional: the processing happens on the issued laptop, no third party receives the file, and there is no new supplier to assess.
- Will it code my data or do thematic analysis?
- No. It produces a transcript and a summary. Coding, memoing, intercoder agreement and the analysis itself belong in MAXQDA, ATLAS.ti, NVivo or Delve, and markdown imports into all four. We are the step before that and have no ambition to be the step after it.
- I need accurate speaker labels for a multi-participant session.
- Speaker identification is a Pro feature and it also runs on your Mac. It saves real time in a two-person interview and is a first pass you should correct in a focus group. If exact attribution across six voices is a requirement of your method, budget for the correction pass rather than assuming the labels are right.
- Half my lab is on Windows or Linux.
- They cannot use it. Ghosty Notes is macOS only and needs a Mac with Apple Silicon; Intel Macs are not supported and there is no Windows or Linux build. On a mixed team this becomes a per-researcher choice rather than a lab standard, which is worth knowing before you plan a study around it.
- Can the whole team see one shared library of interviews?
- No. There is no shared library, no admin view and no team plan. Each researcher's transcripts are files on their own Mac, and sharing means putting those files wherever your project already stores data, under whatever access rules your ethics approval set. Convenient central access is exactly what we do not provide.
Against what research teams currently use
| Ghosty Notes | Alternatives | |
|---|---|---|
| Who receives the recording | Nobody. It is transcribed on your Mac | Otter, Dovetail and Descript receive and store the audio |
| Presence in the session | No attendee. Recording runs off the Mac system audio | A bot joins as a visible participant |
| The Word and Teams route institutions suggest | Needs your own approval, but adds no new recipient | Sanctioned, under an existing institutional contract |
| Analysis and coding | Not offered. Markdown imports into MAXQDA, NVivo or Delve | Dovetail codes and tags inside its own platform |
| Working offline in the field | Yes in local mode, nothing is sent | No, transcription needs the upload |
| Cost per researcher | Free tier, or Pro at $5.00 a month or $54.00 a year | Per-seat subscriptions, research platforms priced per workspace |
| Searching your own corpus from an AI tool | Local MCP server on the free tier, over your own files | Vendor chat, over data the vendor holds |
Questions from research teams
- If I turn on Pro’s hosted AI models, what leaves my Mac?
- The transcript text, and only that. Audio never leaves the device on any plan, and in local mode nothing is sent at all. Hosted models and cross-interview chat send transcript text to that provider, which for participant data is a change to describe in your ethics application before you make it, not after.
- Which Mac do I need?
- A Mac with Apple Silicon. Intel Macs are not supported, because the on-device model needs Apple Silicon to run at a usable speed, and there is no Windows or Linux build.
- How accurate is it on accented speech and technical vocabulary?
- We publish no accuracy figures, because we have no measured data and an invented number would be worse than admitting that. Quality depends heavily on microphone and background noise. Run two real interviews on the free tier and judge it on your own participants and your own recording setup.
- Does anything appear in Zoom, Teams or Meet?
- No. It records the audio your Mac is already playing and hearing, so no attendee is added and no bot appears for a participant to notice or a host to admit.
- Can I get a clean transcript file for an appendix or a repository deposit?
- You get a markdown file per interview, which converts to text, Word or PDF with anything. Anonymising it before deposit is your job; no tool here removes identifiers, and pseudonymising participants is still a manual pass.
- Does it work without an internet connection?
- In local mode, yes. The model downloads once and then runs on the machine, so fieldwork on a bad connection still produces a transcript. Pro's hosted models and calendar sync do need a connection.
- Does it work for in-person and field interviews?
- It depends on what the Mac's microphone picks up: a quiet room with one participant is usually fine, a noisy public setting or a six-person focus group across a table often is not, and no on-device model rescues bad input audio.
- What does the free tier include for a student or a small study?
- On-device transcription, unlimited interviews, the local MCP server and Apple Calendar detection on your machine. There is no minute cap and no account is needed for local use.
- Are there other tools that transcribe on the device?
- Yes. MacWhisper, Superwhisper and Basil all transcribe on-device, so we are not alone in this. The combination here is no bot in the call, markdown in your own folder, and local search through MCP on the free tier.
Keep the participant recording on your own machine
Free for Mac. No bot in the session, no upload of the audio.
Transcribe your next interview locally