Twenty-two interviews that could not leave her laptop

 

A PhD researcher wearing headphones at her desk in the evening, listening to an interview on her laptop with a stack of highlighted printed transcripts

Priya Raman is in the third year of a PhD in sociology at a university in Singapore. Her study is about adult children who look after a parent with dementia at home. Over eight months she sat in twenty-two living rooms and asked people how they get through a week. Most interviews ran an hour. A few ran two, because once people start talking about their mother at three in the morning, you don't stop them.

I'm David, and I work on VoxStudio. Priya is made up, but her problem comes from researchers who have written to us about their interview data. The interviews were the easy part. Everything after them was slow.

The rule in the consent form

Her ethics approval was clear about one thing. Recordings stay on her encrypted laptop and are not sent to an outside service. Participants signed a form that said so. That ruled out the transcription service her lab mates used, and every web tool that wanted an upload.

So for the first nine interviews she typed. One hour of tape took her about six hours, with the foot pedal she borrowed from the department and a lot of rewinding. By December she had nine transcripts, thirteen recordings she hadn't touched, and a supervisor asking when he could see a first pass at the themes.

Then came the meeting that really hurt. Her supervisor had read the nine transcripts and asked a simple question: how many of your participants talked about not sleeping, and can you show me where? She knew at least four had. She found two by memory, spent an evening searching Word documents for "sleep" and "night", and still missed the man who had said only that he "stopped bothering with the bed".

Transcripts that never leave the Mac

She tried the offline transcription guide on one of her own pilot recordings first, the one with a family friend who had agreed to be a practice participant. While still online, she opened Settings, then Local Features, and prepared the English speech model. She waited until the app said it was ready. The guide is careful on one point. If an outside AI provider is configured, the app may try to write a summary, and local processing does not stop that. Priya had never set one up on this laptop. She still followed the guide's advice for data that must not leave the machine: get the models ready, then stay offline for import, review and export.

With Wi-Fi off, she went to Transcribe, then Import files, chose the recording and named the session with the participant code only, "P14 interview 1". She set the input language to English, kept speakers anonymous, left translation off and clicked Transcribe. When it finished, the session showed two badges, "Processing on This Mac" and "Session stored on This Mac". She took a screenshot of both and filed it with her data management notes. Her ethics office liked that more than any promise on a website.

A researcher with a notebook listens to a man talking on his sofa in a Singapore flat, a small audio recorder on the coffee table between them

Speaker 1 is not a person yet

The transcript came back with Speaker 1 and Speaker 2, each block with a time range. The speaker-label guide says plainly that a label is a group of similar voices, not an identity, and that it can be wrong. Short replies and interruptions are the usual misses. Her interviews are full of both. Participants say "yah" and "mm" under her questions all the time.

So she replayed the handovers. She clicked a label, chose Rename, and typed "Interviewer". Then she renamed the other one "P14". The rename applies to every block with that label, so she only did it once per interview. Where a "yah" had landed in the wrong block, she fixed it by hand. She also checked every name of a hospital, a medicine and a day centre against the audio. That was the part that used to eat her week. Now it took about forty minutes for an hour of tape.

Over two weekends she imported the thirteen recordings she hadn't touched, plus the nine she had typed. She wanted the audio and the text together for those too.

Finding the man who stopped bothering with the bed

Then she went back to her supervisor's question. In Knowledge she set the scope to her twenty-two interview sessions and searched by meaning, not by keyword, for people describing being awake at night with their parent. The knowledge base page notes that indexing and search in VoxStudio for Mac run on the Mac. She wasn't sending participants' words anywhere. She left the AI answers alone for this study, because those use whichever service you choose, and her approval didn't cover one.

The results were passages, each with its session, speaker and timestamp. "I sleep on the sofa now so I can hear the door." "Three, four times a night she calls for my late father." And P09, the man she had missed in Word, saying he had stopped bothering with the bed. She opened each passage, read the lines around it, and played the audio. The guide to searching across recordings asks you to do exactly that before you rely on anything, and in interview work tone matters. One "at night" turned out to be about a night shift at work, not caregiving. She left it out.

The answer to her supervisor's question was seven, not four. She counted them herself.

A PhD student points at her laptop while her supervisor leans in to read, a printed spreadsheet with highlighted rows on the desk

Quotes that still point at the tape

Every passage she kept became a saved quote. The quote sheet guide shows the steps. She selected the exact words, clicked Quote, and on the card added a purpose, "Ch 4 night care", and a note about how the participant sounded. VoxStudio keeps the sentences before and after as context, so she didn't have to stretch the selection to include her own question. She used Play from quote to hear each one again, then clicked Verify.

Before her next supervision she opened the Quotes tab in each session, filtered to anything still Pending, and finished the checks. Then Export, Quotes, Excel. Each row had the exact words, the speaker, start and end times, the context before and after, her purpose and note, and whether it was verified. She pasted the rows into the master sheet she uses as her codebook. When her supervisor asked "where's the bit about the sofa?", she gave him a session name and a timestamp instead of a page number in a document he didn't have.

A week later she noticed that P11 had named a day centre she was supposed to anonymise. She corrected the transcript. The quote from that passage was flagged with "Source changed", because a saved quote keeps the text it had when you saved it. She used Compare source, took the corrected text, and the anonymised version went into the sheet.

Her second coder, a postdoc in the same group, only wanted the parts of interview 12 about respite care. Priya exported the transcript as a TXT with speaker labels, turned on Filter rows, typed "respite", and sent him just those rows over the university's file share.

What it takes

VoxStudio for Mac needs an Apple silicon Mac and macOS 15 or later, and you download the local speech models once while online. Check speaker labels and names against the audio, every time. The app says so too. The interview to quote sheet tutorial walks through the whole path with practice files, so you can try it before you touch real participant data.

You get 14 days with every local feature. After that you can keep using VoxStudio on recordings up to 15 minutes long, which is fine for pilots but not for hour-long interviews. For those there is a one-time Mac Lifetime licence, and students get it at half price, which matters on a PhD stipend.

If you have a pile of recordings you promised would stay on your laptop, start with one pilot interview this week and see how long the checking actually takes.

Comments

Popular posts from this blog

570 visitors, 17 clicks, zero downloads: what a too-small OpenAI ad taught me about growing VoxStudio

Speaker Labels vs. Speaker Names: Getting "Who Said What" Right in Mac Transcripts

The quote her editor asked about at 9:40 p.m.