Write Your Novel by Voice: Dictation That Knows Your Characters

MimicReader's Writing Studio transcribes novel chapters by voice using Whisper Large V3 Turbo, with character names piped into Whisper's initial_prompt from your project — so "Ezra" stays "Ezra" instead of becoming "Israel", and "Lyra" stays "Lyra" instead of "Lila". A 3-minute dictation transcribes in ~8 seconds on an RTX 3090 over WireGuard. Audio is stored privately in your Notes for replay; no third-party AI provider (OpenAI, Google, Anthropic) ever sees the recording.

The moment dictation usually fails

I was walking the dog last Tuesday morning, working out a scene in my head. Two characters, Ezra and Lyra, arguing in a half-flooded basement about whether the river had moved on purpose. Good scene. The kind that arrives whole, and disappears the moment you sit down at a keyboard.

So I did what every writer who carries a phone does. I opened a dictation app and started talking. Three minutes later I had the scene. Five minutes later I had the transcript. Every instance of "Ezra" had been turned into "Israel". Every instance of "Lyra" had been turned into "Lila". The basement was now "the bay", and the river had become "the reefer". Half an hour of search-and-replace later, the scene was on the page, but the spark was gone.

This is the universal experience of writing fiction by voice. Generic dictation models have never read your book. They've never even read your character list. They guess at proper nouns from a frequency table of news articles, which is why uncommon names — the exact names fiction tends to use — get rewritten into common ones. Ezra into Israel. Lyra into Lila. Saoirse into "Sersha". Caitlyn into "Caitlin" (or worse).

The fix is structural. The dictation model has to know about your project before it starts transcribing. MimicReader's Writing Studio editor does exactly that.

Why typical dictation fails for novelists

Quick tour of the alternatives, honestly graded:

• Google Web Speech / Chrome dictation — fast, free, runs in the browser. Knows zero about your manuscript. Replaces anything unusual with the nearest common word. Stops listening after ~60 seconds of silence. Audio is sent to Google.
• Apple Dictation (iOS) — slightly better on-device model since iOS 17, but the corpus is still general English. You can add a few "text replacements" but it doesn't bias the model, it just post-processes specific strings.
• Otter.ai — built for meetings. Great at identifying speakers in a Zoom call. Useless at knowing that your protagonist is called Soren and not "soaring".
• Dragon Naturally Speaking — the old king. You can train custom vocabularies, but the workflow is fiddly, the license is hundreds of pounds, and the mobile story is rough.
• Plain Whisper (the open-source model) — the actual transcription engine behind most modern tools. Excellent, but only as good as the prompt you give it. Out of the box, no novel context.

The pattern: none of these tools know your world. They can't, because none of them are inside your manuscript. They're general-purpose tools bolted onto specific work.

How MimicReader's voice dictation works

The Writing Studio editor lives one click away from your character list, your blurb, and the chapter you're currently editing. When you hit the microphone button, here's what happens behind the scenes:

1. Your browser records audio locally (browser MediaRecorder API, no streaming).
2. When you stop recording, the audio blob is uploaded to MimicReader's GPU server over HTTPS.
3. The backend pulls your project's metadata — every named character, the blurb, the working title, key place names.
4. Those names are formatted into Whisper's initial_prompt — a short string of context Whisper uses to bias its decoder. "The following audio is from a novel featuring Ezra, Lyra, Saoirse… set in the half-flooded city of Veridia…"
5. Whisper Large V3 Turbo transcribes the audio on our RTX 3090. The prompt biases it toward your proper nouns — Ezra stays Ezra.
6. The transcript is inserted at the cursor position in your chapter editor.
7. The original audio is saved as a Note in your account so you can re-listen if Whisper misheard a word and you want to fix it from source.

That's the trick. Whisper has supported initial_prompt from day one, but no consumer dictation tool wires it up to a manuscript automatically. We do, because the manuscript is right there.

Your audio stays here

This part matters more than people think. Dictation tools have to do something with your voice. The defaults are not always friendly.

• Google Web Speech sends audio to Google's servers. Subject to whatever they're using it for, and whatever the EU and UK lawyers eventually negotiate about training data.
• Otter, Rev, Trint store your audio on their cloud, indefinitely by default. You can delete it manually. They train on aggregate, deidentified data — terms vary by tier.
• Apple Dictation stays on device for short utterances, but enhanced dictation can phone home.

MimicReader is different by architecture, not by promise:

• Audio is sent to our GPU (an RTX 3090 in a server cabinet outside Glasgow), not to OpenAI, Google, or any third party.
• Transcription runs on a local Whisper model that we host. There's no API key to a third-party transcription service. There's no third party.
• The audio is stored as a Note in your account. Only you can listen to it. You can delete it from the Notes tab at any time. Account deletion (GDPR Article 17) wipes it too.
• We don't train any model on your audio or your manuscript. Full stop.

If you've ever been uneasy about dictating a sensitive scene — a memoir chapter, a confidential interview, the climactic plot twist you don't want a Google ML team to ever see — this is the version that lets you not worry about it.

The mobile workflow

The microphone button works in mobile Chrome and Safari. The actual workflow I use, since the dog won't walk himself:

1. S24 Ultra in pocket, earbuds in, dog leash in left hand.
2. Open mimicreader.ai/app, navigate to the chapter I'm working on.
3. Hit the microphone (a big red dot at the bottom of the editor toolbar — hard to miss).
4. Talk. Whole paragraphs. Whole scenes. Whole chapters if the morning is good.
5. Hit stop. Wait ten seconds. Transcript appears inline.
6. Walk on. Repeat for the next scene.

Whisper Large V3 Turbo is fast. A three-minute dictation transcribes in around eight seconds on our 3090. By the time I've put the phone back in my pocket and watched the dog sniff a lamppost, the text is in the chapter.

There is no battery cost worth mentioning, because no on-device model runs — recording is the only thing your phone does, and recording uses almost nothing. There is some network cost: a three-minute Opus blob is around 1 MB. Cheap on 4G, free on home Wi-Fi.

What gets saved (and why)

Two things, every time you finish a dictation:

1. The transcript — inserted at the cursor, instantly editable. This is the primary output.
2. The original audio — archived as a Note in your account, with the timestamp of the recording and a link back to the chapter it was dictated into.

The audio archive matters because Whisper, even with bias, isn't perfect. Maybe it heard "the morning when I" as "the moaning weather". Maybe the wind caught one word and turned it into nonsense. With the original audio one tap away in Notes, you can replay the exact phrase, hear what you actually said, and fix the transcript. No reconstructing from memory three days later.

The Notes view also lets you stack up raw audio for later. Dictate four scene sketches on Monday's walk. On Wednesday evening, sit down with a coffee, open the Notes tab, listen back, and decide which ones are worth keeping. It's the same pattern the voice-notes-to-audiobook pipeline uses, just bound directly to the editor instead of to a Promote-to-chapter button.

How to enable it

Writing Studio is currently behind a feature flag, because parts of the editor are still being polished. To turn it on:

1. Open /settings in the MimicReader app.
2. Scroll to Writing Studio.
3. Flip the toggle on. Save.
4. Open any project chapter. The editor toolbar now has a microphone icon next to the formatting buttons.
5. Click it. Allow microphone access when the browser asks. Speak.

That's it. No extension, no install, no separate app. The microphone button is part of the editor the moment Writing Studio is enabled.

Try the workflow

If you've ever opened a fresh document at 7am with a scene in your head and lost it to the first paragraph of typing, this is the workflow that fixes that. Voice keeps the spark. Project-aware dictation keeps your characters' names.

Where to go next

• The full 4-step pipeline from voice notes to a published audiobook
• Why we built writing and audiobook generation in one tool