For UX, qualitative & academic researchers

Interview transcription for researchers that stays on your Mac.

Quietly transcribes on your Mac and writes plain Markdown notes to a folder you pick. Transcription sends no audio or text to OpenAI, Otter, or anyone else, which matches what your consent form promises.

Download for Mac For your IRB
Free 7-day trial On-device Whisper Local AI chat 99 languages $49 once
What you get
Interview folders, anonymisable.

Plain .md with participant ID, date, and duration in the front-matter. Atlas.ti, Dedoose, and NVivo all read it.

Local · No cloud
P07-2026-04-22.md 58 min · KR
P08-2026-04-23.md 42 min · EN
P09-2026-04-25.md Transcribing
The friction

Cloud transcription tools and IRB protocols don’t mix.

The Whisper API and Otter ship audio off-device. Both send it to OpenAI’s or Otter’s servers, which blocks IRB approval on most studies that touch identifiable speech.

Rev human transcription costs $1.50 a minute. It is accurate, but a 60-minute interview costs $90, and your audio still leaves your laptop.

Self-hosting Whisper costs you a weekend. You end up with half-named files, no metadata, no batching, and no front-matter. Then participant 14 sends an .m4a and the pipeline breaks. (The DIY commands are in how on-device Whisper transcription works on a Mac, if you want to try it.)

Dovetail and Reduct charge cloud-tier prices to host data your IRB already cleared you to keep local. You pay to add risk.

Why Quietly fits

The boring local-first pipeline you wished existed.

On-device

Audio + transcript stay on your Mac.

Whisper large-v3 transcribes on your Mac, and Quietly deletes the audio file unless you keep it. Nothing leaves the device unless you point Chat at a cloud model.

Markdown

Open in Atlas.ti, Dedoose, or NVivo.

Front-matter has participant ID, date, and duration. The body is clean, LLM-polished Whisper output. Easy to anonymise, import, and cite.

Any audio file

Transcribe existing recordings, no CLI.

Pick Transcribe a file… and choose any common audio or video file (.m4a, .wav, .mp3, .mp4, …). Each becomes a Markdown note. No Python env, no GPU drivers.

99 languages

Korean & English, side by side.

Whisper detects the language per recording. Code-switching interviews (KR + EN) stay coherent, since large-v3 handles them well. We test it daily on Korean academic interviews.

In-person

Lab interviews, fieldwork, intercepts.

Mic-only mode covers in-person sessions. Open Quietly and press ⌘R to start a recording, with no full-screen app and no bot in the room.

Local AI chat

Ask your interviews on a local model.

Point Chat at a local model in Ollama or LM Studio on your Mac and ask for themes, quotes, or a summary. The transcript stays on the device.

Lab meetings and one-on-ones

You say unpublished ideas out loud first.

Before an idea reaches a preprint, it comes up in a lab meeting, an advisor one-on-one, or a call with a collaborator. A cloud note taker uploads that audio. In September 2026, mathematicians alleged that unpublished work typed into AI products may have reached a model, which OpenAI denies. We cover what the dispute means for your lab in is it safe to share unpublished research with AI?

Quietly records these conversations with no bot in the call and keeps the audio and transcript on your Mac. Run Chat on a local model in Ollama or LM Studio and you can ask what the group decided without the transcript leaving the machine. For meetings with more than two people, Quietly splits the transcript into speaker turns today. Per-speaker labels are coming.

Check it yourself

Test it the way Northwestern suggests.

Northwestern IT lists Whisper-based apps that run on your computer for research interviews. It tells you to “Turn off all internet connections from your device to check that the software is working locally on your computer and not sending any data to the cloud.” It also says: “Do not use AI transcription software that sends data to third-party services or requires you to provide an API key in order to use the software.” Quietly’s transcription does neither, and neither does Chat when it runs on a local model in Ollama or LM Studio. Only a cloud model (OpenAI or Anthropic with your own key) would cross that line. You can run the offline test in five minutes:

  1. Install Quietly, let it download its models once, and pull a model in Ollama or LM Studio.
  2. Turn off Wi-Fi and unplug Ethernet.
  3. Transcribe an interview file, or press ⌘R and record a test conversation.
  4. Watch the Markdown note appear with no connection.
  5. Point Chat at a local model in Ollama or LM Studio and ask a question about the note, still offline.

Quietly goes online for model downloads, update checks, license validation, Google Calendar if you connect it, and Chat if you pick a cloud model. For what other universities allow, see what 12 universities say about Otter.ai.

For your IRB

The sentence you can paste straight into your protocol.

It reads like the privacy clause your IRB already accepts.

Audio recordings will be transcribed using OpenAI Whisper running locally on the researcher’s device (Quietly, meetquietly.com). No audio or transcript will be transmitted to a third-party server. Any AI-assisted analysis in the software’s chat feature will use a local model (Ollama or LM Studio) on the same device; no cloud AI provider will be used. Recordings will be deleted after transcription; anonymised transcripts will be retained for analysis.

We can provide a one-page technical brief for your IRB. Email support@meetquietly.com.

Your analysis stack

Plain Markdown, so your QDA tools read it.

Atlas.ti
Import as plain text. Quote-level coding works on the transcript paragraphs as they are.
.md / .txt
Dedoose
Upload the Markdown. Dedoose treats each file as a document for thematic coding. It runs in the cloud, so check your protocol first.
.md
NVivo
Convert front-matter to NVivo classifications with a tiny script. Examples on request.
.docx ←
VS Code · ripgrep
Tag, search, and pull quotes across the whole corpus. It’s plain Markdown.
folder
Pandoc / LaTeX
Convert quotes into your dissertation’s bibliography format with one script.
cli
FAQ

Things researchers ask first.

Is the Whisper model running on my laptop?
Yes. Quietly downloads Whisper large-v3 once, then runs it on your Mac's CPU and GPU. Quietly processes the audio file on disk and never sends it over the network. The app only goes online for one-time model downloads, update checks, license validation, Google Calendar polling if you connect a Google account, and Chat if you point it at OpenAI or Anthropic instead of a local model in Ollama or LM Studio. With a local model, Chat runs on your Mac and needs no connection. We walk through the transcription pipeline in how on-device Whisper transcription works on a Mac.
How accurate is local Whisper vs. Rev or human transcription?
On clean recordings (interview-grade mic, quiet room), Whisper large-v3 is comparable to professional human transcription in English and strong in Korean. An on-device LLM cleanup pass then fixes fillers, false starts, and punctuation, so the transcript reads as clean as a cloud service's, and it never leaves your Mac for cleanup. Noisy fieldwork audio will need a verification pass, but at $49 instead of $90 an hour, the math still works.
Can I diarize speakers?
Today Quietly splits the transcript into speaker turns, with a new block each time the speaker changes, which works well for two-person interviews. Full per-speaker diarization (labeling each participant, robust enough for focus groups of four or more) is coming soon, along with per-channel diarization for when you record each participant on a separate track.
What happens to the audio file?
By default, Quietly deletes the audio once transcription completes. You can keep it, per recording or always. It stays on your disk in the same folder, with a filename that matches the .md note.
Will my IRB ask for a SOC 2 report?
Some will. Quietly is local-first software, so there is no service to audit. We provide a one-page technical brief that covers data flow, network calls, and where Whisper runs, and most IRBs accept it. For the strict ones, the brief states that audio and transcripts never cross the network while Chat runs on a local model or stays off, and that the app only goes online for model downloads, update checks, license validation, and optional Google Calendar polling.
Can I use this for clinical / IRB-restricted work?
Several research groups use Quietly for IRB-approved interview studies. We don't market Quietly as HIPAA-certified, since that is a service-level claim and Quietly is a local app. The data handling does match what most IRB protocols require: on-device processing, no third-party transmission, and researcher-controlled retention, as long as Chat stays off or runs on a local model.
Does the built-in Chat send my transcripts anywhere?
Only if you pick a cloud model. Chat can run on a local model through Ollama or LM Studio on the same Mac, so your questions and the transcript never leave the device and no API key is involved. If you connect your own OpenAI or Anthropic key instead, Chat sends the meeting you ask about (transcript, notes, and details such as attendees) to that provider. Both exclude API data from model training and keep it for up to 30 days by default. For IRB-restricted data, use a local model.

Replace your transcription budget with $49.

Free 7-day trial · No account, no card. 7-day refund, macOS 13+. Approve once, use it for the whole dissertation.

$ 49 once · 2 Macs
Download for Mac

Less than 1 hour of human transcription