The witness said "it". Her notes said "her".

By David

An HR manager interviews a warehouse worker in a small meeting room, a handheld audio recorder on the table between them


Nadia Rahman is the HR business partner for a cold-chain logistics company in Tuas. About six hundred people work there, most of them on shifts in the warehouses. Most days are hiring paperwork and shift-roster arguments. A few times a year there is a grievance investigation, and those are the weeks she dreads.

I'm David, and I work on VoxStudio. Nadia is made up, but her week is built from what people who handle confidential interviews for a living have told us. They have to get every word right, and they are not allowed to send the recording anywhere to help them do it.

The appeal she still thinks about

Last year Nadia ran an investigation into a complaint on the night shift. She interviewed five witnesses in a small meeting room with a notepad, writing as fast as she could. Each interview took about forty minutes. Typing up her notes took another three hours each, mostly at home in the evening.

At the appeal, one witness read his statement and said, "I never said that." Her notes had him saying a supervisor "pushed her". He said he had told her the supervisor "pushed it", meaning a pallet trolley. There was no recording. It was her handwriting against his memory, and the panel sent part of the case back.

After that, she started recording interviews, with consent, on a company-issued handheld recorder. That solved half the problem. The other half was what to do with the audio. The company's legal counsel had one firm rule. Investigation material stays on company devices. It does not go to an online transcription service or a chatbot.

So the recordings sat on her laptop, and she was still typing statements by hand while playing the audio back.

A complaint on a Tuesday

On a Tuesday in November, a picker on the chiller floor filed a grievance against her shift lead. She said he had shouted at her in front of the team and shoved a pallet trolley into her leg. Four people had been nearby. Nadia booked the meeting room for all of Wednesday.

This time she had VoxStudio for Mac on her work MacBook. A month earlier, with the laptop still online, she and the IT administrator had opened Settings, then Local Features, and prepared the English speech models until the app showed them as ready. Then they read the offline transcription guide together, because it says something they needed to know. If an external AI model is configured, the current version tries to write a summary automatically, and the local badges do not stop that request. Their setup had no outside AI service configured. Nadia still decided to switch off Wi-Fi before she imported anything from a case, as the guide suggests when no transcript content may leave the Mac.

Wednesday evening, with the Wi-Fi off

On Wednesday, each interview started with her reading out why she was recording and who would hear it, and the witness saying they agreed. That evening she copied the four files from the recorder to her Mac, turned off Wi-Fi and opened Transcribe, then Import files, then Choose media. She titled each session with the case number and a witness letter, never a name, for example "GR-0412 Witness C 12 Nov". She set the input language to English, left Speaker count on Auto-detect and left translation off.

Weeks before, she had saved her own voice to VoxStudio's voice library from a practice recording of herself. Under Recognize people she ticked only herself. VoxStudio names a voice only when it clearly matches a person you selected, so her questions came out under "Nadia" and each witness came out as Speaker 2. She did not save any witness's voice.

When the first transcript finished, she checked the two labels on the session, "Processing on This Mac" and "Session stored on This Mac". The guide is clear that one badge is not proof of the other, so she looked for both. Then she clicked Speaker 2, chose Rename Speaker 2, and typed "Witness C". The change applied to every block in the transcript.

The line that mattered

Witness C was a forklift driver who had been about four metres away. Halfway through the interview, the transcript read, "Then he pushed her, near the leg, and she shouted back."

It was exactly what Nadia had expected to hear, so she clicked Play from this segment and listened with her earphones in.


The same HR manager at her desk in the evening, earphones in, listening back to an interview on her laptop with a handheld recorder beside it

He had said, "Then he pushed it, near the leg." The "it" was the trolley. His next sentence, which the transcript had right, explained that the shift lead had shoved the trolley away from the racking and it had rolled into the picker's shin. That is still a serious finding. But it is not the same as saying the shift lead put his hands on her.

She chose Edit text, changed "her" to "it", and clicked Done. VoxStudio keeps the original output, so she could show later what the software had produced and what she had corrected by ear.

Then she went through the short answers. The speaker-label guide warns that one-word replies and interruptions are the ones most often labelled wrong. Two of Witness C's "yes"es had landed under her name. She fixed those too. Reviewing a thirty-eight minute interview took her about fifty minutes. Typing one used to take most of an evening.

More on speaker labels in interview transcripts.

A statement the witness could read

On Thursday morning she opened Export above the player and chose Transcript, TXT, Original, and With speaker labels. She printed the file, which has each block under the speaker's name and no time codes. Witness C read his statement at the meeting room table, corrected the spelling of a colleague's name in pen and initialled every page.

If anyone asks at an appeal what he said, she has the recording, her corrected transcript and his initials.


A warehouse worker initials a printed statement at a meeting room table while the HR manager sits across from him

Writing the report by talking

The report was the part she used to put off. She writes it in Word on the same laptop, with a summary of each statement and a finding on each allegation.

In VoxStudio Settings, under General and Dictation, she had set Shortcut action to "Dictate into the current app" and allowed Accessibility access. Now she clicks into the Word document and holds Control + Option + Space. She speaks a sentence or two and lets go. VoxStudio recognizes the speech on the Mac and pastes the text at her cursor. For a longer paragraph she taps the shortcut once, talks it through and presses it again to insert. According to the dictation guide, the audio goes to a temporary file, is recognized locally and is discarded once the text is in.

She still edits every paragraph and checks names and dates, but now she dictates from a transcript she has already checked. The draft was ready by Monday afternoon. Last year's report had taken her until the end of the following week.

What it does not do

VoxStudio does not decide whether you are allowed to record someone. Her company's counsel decided that, and Nadia gets consent at the start of every interview. Local transcription also does not make AI features offline. If she asked VoxStudio's Knowledge for an answer or turned on summaries with an outside model, the text that task needs would go to that service. For case files she doesn't use those features. Our post on where AI meeting notes go on a Mac explains how the paths differ.

Sign-in and access checks need the internet. A Lifetime authorization stays valid offline for up to 14 days, then refreshes online.

If your recordings can't leave the room

VoxStudio for Mac needs Apple silicon and macOS 15 or later. You get 14 days with every local feature, and after that you can keep using it on recordings up to 15 minutes long. Mac Lifetime is a one-time $49. Students get it at half price, and journalists with press credentials get 30% off. The pricing page has the details.

If you want to test it the way Nadia did, use a short recording you are allowed to use, turn off Wi-Fi, and check every line that someone might one day dispute.

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.