Voice Tools for Journalists: Interview to Draft, Without Compromising Sources

The Journalist's Two Voice Problems

Reporting is a voice-heavy job. You talk to people, then you write about what they said. Two voice-driven tasks occupy most of the work:

Interview transcription. You have a 45-minute recording. You need a transcript to quote from, find threads you missed, and verify attribution. Typing it up by hand takes 3–4x the recording length. A working reporter might spend 15 hours a week on this.

Drafting. Once you have quotes and structure, you write. Writing is typing, for most people. Typing is slower than thinking, and much slower than speaking. For long reported features, draft speed is the ceiling on output.

Cloud services like Otter, Rev, Descript, and Trint have attacked both problems. They're good products. They solve the speed problem. They create a different one.

The Source Protection Problem With Cloud Transcription

When you upload an interview recording to a cloud transcription service, you're doing three things you may not have thought through:

You're creating a record on a third-party server. Your source's voice, your source's words, your source's possibly-confidential statements — stored on infrastructure you don't control, under a retention policy you may not have read, potentially accessible to the vendor's employees, security contractors, and (in some jurisdictions) their government.

You're creating something subpoena-able. Shield laws protect journalists' confidential sources, but they protect the journalist, not necessarily the vendor. A prosecutor or litigant seeking the identity of a source can subpoena the vendor directly. Some vendors have fought subpoenas; others have complied.

You're expanding your breach surface. Every major transcription vendor has faced security incidents. If a vendor gets breached while holding your audio, your source's identity could be exposed through voice recognition, background sounds, metadata, or the content of what they said.

The Local Transcription Alternative

OpenAI released Whisper as an open-source speech recognition model in 2022. It's genuinely state-of-the-art — competitive with or better than most commercial transcription services. And critically for journalism: it can run entirely on your Mac, with no network connection required.

Apple Silicon (M1 and newer) can run the largest Whisper model locally at faster-than-real-time speeds. A 45-minute interview transcribes in 8–12 minutes on an M2 MacBook Air, without any audio ever leaving the device.

EmberType is a Mac dictation app that wraps Whisper in a usable interface. For journalists, the relevant capabilities:

• Transcribe existing audio files — drop in a Zoom recording, phone interview, or voice memo and get a transcript back in minutes
• Transcribe live with system audio capture — for remote interviews, capture the call audio without a bot joining (covered in our Zoom piece)
• Voice dictation into any Mac app — type drafts by speaking directly into Google Docs, Word, Scrivener, Ulysses, iA Writer, or your CMS
• Custom dictionary — add source names, place names, agency acronyms, and terminology specific to your beat
• Multilingual — transcribe Spanish, Arabic, Mandarin, French, and more, all locally
• No cloud, ever — the model runs on your Mac's Neural Engine

The Interview Transcription Workflow

1. Record the interview well

Garbage in, garbage out. Whisper is excellent but not magic. Use a decent microphone for in-person interviews (a USB condenser or shotgun mic is dramatically better than built-in laptop audio). For phone interviews, record through a tool that captures both sides cleanly — TapeACall for iPhone, or a desktop capture via EmberType's system-audio feature.

2. Get the file onto your Mac

AirDrop for iPhone recordings, direct download from Zoom/Riverside, or just a file in your Downloads folder. No cloud sync required.

3. Transcribe locally

In EmberType, select the audio file and choose your Whisper model. For interview transcription, use Large v3 — the most accurate model. Transcription runs on the Neural Engine; you can keep working on other things while it runs. For a 60-minute interview, expect 10–15 minutes of processing on an M2 or newer.

4. Clean up with AI Enhancement (optional)

Raw Whisper output is a wall of text. EmberType's AI Enhancement feature can post-process the transcript: add speaker labels (based on pause patterns), break into paragraphs, remove excessive filler words, or produce a summary. All of this also runs locally if you use a local LLM, or calls your API of choice if you prefer.

5. Import into your writing tool

The transcript is yours. Paste into Scrivener as a research note, a Google Doc for highlighting quotes, an Obsidian vault for permanent reference, a DEVONthink database, or whatever system holds your reported material.

The Draft Dictation Workflow

Most journalists who try voice dictation for drafting have the same reaction: “This is weird, I hate it, wait, I finished the draft in 40 minutes.”

Here's why it works and how to make it work for you.

Dictate the bad first draft

Anne Lamott's “shitty first drafts” advice is perfect for voice. The point of the first pass is structure, arc, and voice — not polish. Speaking gets you there faster because you're not correcting typos, reaching for synonyms, or tweaking commas. You're just saying the thing.

Open your doc, press EmberType's shortcut, and narrate the piece. Start with the lede you've had in your head for two days. Move through your reporting. Don't stop to fix things. When you hit a spot where you don't have language yet, say “brackets TK placeholder about the funding history brackets” and keep moving.

Edit with the keyboard

Dictation is for generation. Editing is for the keyboard, where precision matters. Once you have a full draft, go back and edit conventionally. This is where the craft happens; voice is just what gets you to a complete artifact fast enough that editing becomes the main task.

Use voice for long-form, keyboard for short-form

For a 200-word news hit, voice has no advantage. For a 2,500-word feature, voice is transformative. The break-even is around 400–600 words, depending on your typing speed. We tested the speed difference directly in Voice Dictation vs Typing at 90 WPM.

Dictate notes too

Voice isn't just for drafts. Dictate interview notes, reporting observations, scene descriptions you want to remember, questions you need to ask. Capture more of what's in your head because the friction is lower. The research phase of a story benefits almost as much as the drafting phase.

Specific Tools That Fit the Journalism Workflow

EmberType is a horizontal tool — it works in any Mac app. For journalists specifically, here are the downstream apps where dictation pairs best:

• Google Docs — collaborative drafts, shared with editors; dictation works directly in the document pane
• Scrivener — long-form reporting projects; dictate into any text field including research notes and outlines
• Ulysses / iA Writer — minimalist drafting environments; dictation works in the main editor
• Notion — reporting workflow, source tracker, beat notes (see our Notion piece)
• Word — for outlets that still prefer it; works in the main compose area
• CMS web interfaces — WordPress, Ghost, Substack, newsroom CMSes; dictate directly into the browser editor
• Signal, ProtonMail — for source communication; dictation keeps content off your clipboard

Multilingual Reporting

This is underrated. Whisper's multilingual support is excellent and it's all local. A journalist covering Mexico can transcribe Spanish-language interviews without uploading them to a vendor. A reporter covering Hong Kong can handle Cantonese locally. A foreign correspondent can produce transcripts in the country's language before translation, without depending on in-country internet or vendor availability.

Languages EmberType supports well: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Mandarin, Cantonese, Japanese, Korean, Arabic, Hindi, Turkish, Polish, and dozens more. Accuracy varies by language (English is the best-trained), but the privacy property is identical for every language: nothing leaves the Mac.

When You Should Still Use a Cloud Service

Not every interview is sensitive. For a public-official press conference that's already being broadcast live, the marginal privacy gain of local transcription over cloud is zero. If your newsroom standardizes on a cloud service for workflow reasons — shared transcripts, editor collaboration, integrated timestamping — that's a legitimate choice.

The argument for local isn't “never use cloud transcription.” It's: know the difference, and for confidential-source material, default to local. A good reporter has both tools on their Mac and picks the right one per story.

Bottom Line

Voice work is the heart of reporting. The tools that accelerate it used to require trading source privacy for speed. That trade is no longer necessary on Mac. Whisper Large v3 runs locally on Apple Silicon at pro-grade accuracy. EmberType wraps it in a shortcut-driven interface that handles both interview transcription and draft dictation.

For any journalist working on sensitive beats — investigations, national security, sources under NDA, whistleblowers, vulnerable communities — local transcription should be the default and cloud the exception, not the other way around.

---