Dictation for Academics: Drafting Papers, Grants, and Reviews by Voice

Academic dictation works best when the writing is shaped like prose. Argument, interpretation, significance, limitations, reviewer critique, and grant narrative are natural candidates for voice. Notation, references, and structural markup are usually faster to handle with a keyboard.

The practical workflow is a two-pass method: speak the argument first, then add citations, equations, and precise formatting at the keyboard. This guide shows how to use citation placeholders, LaTeX snippets, discipline-specific vocabulary, and a careful data-handling checklist without interrupting the drafting flow.

Why Your Methods Section Fights Back When You Dictate It

Scholarly prose is not uniformly dense. A discussion paragraph is mostly ordinary English with careful hedging. A methods paragraph is closer to source code: subscripts, units, symbol definitions, instrument identifiers, and parenthetical references packed into every sentence. Speech encodes the first well and the second badly, because the acoustic signal carries no information about whether "sigma sub i" is a subscript, a footnote marker, or a variable you defined three pages earlier.

The failure mode is not transcription error. It is repair work. You say "n equals forty-two comma p less than point zero one" and you get a sentence you now have to reformat, which breaks the momentum that made dictation worth using in the first place. Every symbol you speak is a small debt you pay back with the keyboard.

So sort your writing by shape before you decide what to dictate. Argument-shaped text has clauses, transitions, and qualifiers. Apparatus-shaped text has symbols, keys, and cross-references. Dictate the first category and type the second.

Treat dictation as a drafting tool for a specific text shape, not as a universal input method. Choosing the right sections before you start prevents most of the frustrating repair work.

The Two-Pass Method: Voice for Argument, Keyboard for Apparatus

Pass one is a talking pass. You speak the argument spine at conversational speed with no formatting, no citations, no equations, and no going back to fix wording. The output is plain text with placeholders where the apparatus will go. The only rule is that you do not stop moving.

Pass two is a keyboard pass. You go back through the draft and insert references, equations, figure callouts, cross-references, and the precise hedging that scholarly writing requires ("suggests" versus "demonstrates" is a decision, not a transcription). This pass is slower per word but it is bounded, and you are editing rather than generating.

Separating them protects the one thing voice input is genuinely good at, which is sustained narrative momentum. When you interleave drafting and formatting, you interrupt your own argument every twenty seconds. The two-pass split moves all interruptions into a later block where they cost nothing.

Use this repeatable cycle for a manuscript section:

• Outline by keyboard. Write one claim-focused bullet for each planned paragraph.
• Talking pass. Speak each bullet into a full paragraph, say "cite" placeholders out loud, and keep moving forward.
• Cooling period. Step away before editing so you can hear where the draft sounds more confident than the evidence supports.
• Keyboard pass. Resolve placeholders, insert equations, tighten hedges, and fix paragraph order.
• Read-aloud check. Read the final version aloud to catch awkward seams between spoken and edited sentences.

If you want the general version of this loop before adapting it to academic writing, our dictation workflow guide for writers covers the drafting and editing split in more detail, and the voice-first macOS setup article covers window and focus management so the talking pass does not get derailed by notifications.

Speaking Citations Without Breaking Flow

Do not dictate real citation keys mid-sentence. "Open paren Smith and Alvarez comma twenty twenty five comma p one four two close paren" is four seconds of speech that produces a string you will retype anyway, and it derails the sentence you were building.

Use a spoken placeholder instead. Say "cite Smith twenty twenty five" and let the transcript contain something like `cite Smith 2025`. Then one normalization pass converts every placeholder into a real key. The convention matters more than the exact format: pick one, use it every time, and make it something that never appears in normal prose.

```bash

# Normalize spoken placeholders into Pandoc citation keys

# "cite Smith 2025" -> "[@smith2025]"

sed -E 's/cite ([A-Z][a-z]+) ([0-9]{4})/[@\L\1\E\2]/g' draft.md > draft-cited.md

# Multi-author: "cite Smith and Alvarez 2025" -> "[@smithAlvarez2025]"

sed -E 's/cite ([A-Z][a-z]+) and ([A-Z][a-z]+) ([0-9]{4})/[@\L\1\E\u\2\3]/g' \

draft-cited.md > draft-final.md

```

This lines up cleanly with Zotero if you use Better BibTeX, which generates stable citation keys in the `authorYear` shape that Pandoc and BibLaTeX both understand ^[1]. Pandoc then resolves `[@smith2025]` against your exported `.bib` file at render time ^[2]. If your keys use a different pattern, adjust the regex once and the convention holds for every future draft.

Verbal markers for the rest of the apparatus, using the same principle of short, unmistakable tokens:

• "figure ref one" for a figure cross-reference, converted later to `\ref{fig:one}` or your editor's equivalent
• "table ref two" for tables, same conversion
• "section ref methods" for internal section pointers
• "footnote open" / "footnote close" to bracket footnote text you will move later
• "todo check" for anything you need to verify, so a single search finds every open question
• "quote open" / "quote close" for direct quotations, which need exact wording checked against the source anyway

LaTeX and Math: What to Dictate and What to Never Dictate

The rule of thumb: dictate symbol names and short inline expressions, type anything with nested fractions, subscripts, superscripts, matrices, or aligned environments. Speech is linear and math is a tree. Once the tree has depth, spoken input costs more than it saves.

Short expressions are workable when the repair is trivial. Here is a representative before-and-after example:

```text

Spoken transcript:

"The estimator converges when lambda is greater than zero and n

approaches infinity, which follows from equation ref three."

Target LaTeX:

The estimator converges when $\lambda > 0$ and $n \to \infty$,

which follows from Eq.~\ref{eq:convergence}.

```

Two symbol substitutions and one reference fix. Now compare that with a nested case:

```text

Spoken transcript:

"sigma hat squared equals one over n minus one times the sum from

i equals one to n of open paren x sub i minus x bar close paren squared"

Target LaTeX:

\hat{\sigma}^2 = \frac{1}{n-1}\sum_{i=1}^{n}\left(x_i - \bar{x}\right)^2

```

You spoke twenty-eight words to avoid typing one line. Type the line.

Snippet expansion for the symbols you use constantly

The middle ground is snippet expansion triggered by short spoken tokens. Espanso is cross-platform and configuration-driven, which makes it easy to keep a per-project snippet set in version control ^[3]. macOS text replacement and Raycast snippets both work too if you prefer fewer moving parts.

```yaml

# espanso: ~/.config/espanso/match/latex.yml

matches:

- trigger: "symlam"

replace: "$\\lambda$"

- trigger: "symsig"

replace: "$\\sigma$"

- trigger: "eqref"

replace: "Eq.~\\ref{eq:}"

- trigger: "cintv"

replace: "95\\% confidence interval"

```

The other habit that helps: scaffold your environments before you talk. Build the `\begin{itemize}`, `\begin{theorem}`, or `\begin{align}` blocks by keyboard, then dictate the prose inside them. Speaking structural commands out loud is where LaTeX dictation usually falls apart, because backslashes and braces have no natural spoken form. For Overleaf specifically, dictate into a local plain-text scratch file and paste in, since browser editors can behave unpredictably with system-level text insertion.

Custom Vocabulary Beats a Bigger Model for Subfield Jargon

The words that break academic dictation are predictable: Latin binomials, author surnames, instrument model names, gene symbols, reagent catalog numbers, and acronym strings your subfield uses fifty times a paper. These are rare tokens. A larger general model has seen them barely more often than a smaller one, which is why upgrading model size often fails to fix the errors that annoy you most.

A custom vocabulary list is often the more useful move. Build a focused list from two sources you already have: the abstracts in your reference library and your recent manuscripts. Export the abstracts, count word frequencies, and remove ordinary dictionary words to reveal the terminology specific to your field.

Test the vocabulary rather than judging it by feel. Pick a short, fixed paragraph from one of your own published papers, read it aloud, and count repair edits. Repeat the same reading after adding the vocabulary file. The difference shows whether the list helps your voice and your subject area. If you want the background on how these error counts are computed, our word error rate explainer covers what the metric does and does not capture.

The practical choice for academics comes down to language coverage. An English-first model like Parakeet is attractive if you write and speak only in English ^[4]. A multilingual Whisper variant matters if your drafts contain German philosophical terms, Spanish interview quotes, or transliterated fieldwork vocabulary, because an English-only model will force those into the nearest English-sounding word every time.

Peer Reviews Are the Best Place to Start

Reviews are the ideal first dictation project because they are argument-shaped, deadline-driven, and formatting-free. Nobody expects a review to contain equations. There are no citations to resolve beyond the occasional "see Smith 2025 on this point." The whole document is you explaining what you think and why, which is exactly what speech is good at.

Use a spoken template so you never stall on structure:

1. 1.Summary. Two to four sentences restating the paper's claim and contribution in your own words.
2. 2.Significance judgment. One paragraph on whether the contribution warrants publication in this venue.
3. 3.Major concerns, numbered. Each one a stand-alone paragraph.
4. 4.Minor concerns, numbered. Typos, unclear figures, missing details.
5. 5.Confidential comments to the editor. Short, direct, separate from the author-facing text.

For a long review, split the work into separate talking passes: one for the summary and major concerns, then another for minor points after you revisit the figures. This makes it easier to maintain the same level of detail from beginning to end.

One caution about tone. Spoken criticism carries an edge that survives transcription, and "I have no idea what the authors think they are measuring here" sounds far worse in text than it did in your office. Read every dictated review once with a single question in mind: would I be comfortable if the author knew this came from me? Then soften anything that fails.

Also check your funder and journal policies before any AI tool touches review material. NIH's standing notice prohibits generative AI use in the peer review process ^[7], and COPE's position on AI in decision-making sets out similar expectations about reviewer responsibility and confidentiality ^[8]. Transcribing your own spoken words is a different activity from generating content, but the policy landscape is specific and worth reading rather than assuming.

The Confidentiality Problem Nobody Puts in the Ethics Statement

Almost everything an academic dictates is confidential. Manuscripts under review, embargoed results, unfunded grant ideas, human-subjects details, tenure letters, and admissions discussions all fall under agreements you signed or policies your institution enforces. When you dictate that material, you create an audio recording of it.

For confidential work, start with four questions:

• Processing location
• Retention terms
• Model-improvement terms
• Access policy

Compare the answers with your journal agreement, research protocol, and institutional policy before choosing a workflow.

There is also a consent dimension worth separating out. Push-to-talk dictation captures only your own voice, which is a meaningfully different situation from meeting transcription tools that record other people. That distinction is why dictation and meeting capture deserve separate policy treatment, a point we covered in the voice data audit walkthrough.

Choose whether an on-device workflow fits your policy, then map the full path from microphone to editor, backups, and downstream apps. Our local speech recognition guide walks through that review.

Your First Week: A Concrete Academic Dictation Plan

Work through this in order. Each step depends on the one before it.

• Setup. Install your dictation tool, choose a push-to-talk key you can reach without looking, and test the microphone where you usually write.
• Vocabulary build. Export abstracts from your reference manager, identify recurring non-dictionary terms, and add co-author surnames and instrument names.
• Baseline measurement. Read a fixed paragraph from one of your own published papers and count the repair edits.
• First peer review. Dictate a review using the spoken template and wait until the talking pass is complete before editing.
• Discussion section. Outline claim-focused bullets, then dictate one paragraph per bullet before adding citations.
• Keyboard pass. Resolve placeholders, insert equations, and tighten hedging.
• Re-measure. Read the same paragraph again and compare the repair edits with your baseline.

Track repair edits for the same sample paragraph each week. It gives you a useful view of vocabulary coverage, microphone quality, and speaking clarity without relying on a generic leaderboard score ^[5].

FAQ

Can I dictate equations? Short inline expressions, yes: symbol names, comparisons, and simple relations convert with one or two substitutions. Nested fractions, matrices, and aligned environments, no. Type those and dictate the prose around them.

Does it work for non-English quotations? Use a multilingual model when your draft includes more than one language. Whisper large-v3 publishes multilingual coverage in its model card ^[6], while English-first models are tuned for English speech. If your work is bilingual, test the languages you actually use before choosing a model.

What about co-authored drafts? Dictate into a local scratch file, run your placeholder normalization, then paste into the shared document. Live dictation directly into a collaborative editor while a co-author is typing creates merge noise nobody wants.

Will reviewers notice a dictated draft? They notice under-edited drafts, not dictated ones. Spoken prose tends toward long sentences, redundant transitions, and unearned confidence. The read-aloud check plus one editing pass removes those signals.

Is dictation allowed under journal and funder policy? Read the specific policy. NIH prohibits generative AI in peer review ^[7], and COPE's guidance places responsibility on the reviewer for confidentiality ^[8]. Transcribing your own speech locally is different from generating text, but you should be able to explain that distinction if asked.

Try this with your own writing

Export abstracts from your reference manager, identify the recurring terms that a standard dictionary would not know, and load them as a custom vocabulary. Then read a paragraph aloud from a paper you already published and count the repair edits. That gives you a baseline you can compare against later.

Keep typing the notation-heavy sections. Use voice for discussion, review, and grant narrative, where developing the argument matters more than entering symbols. Once you separate those jobs, dictation becomes a practical academic writing tool instead of another formatting problem.

References

[1] Better BibTeX for Zotero, Citation Keys. https://retorque.re/zotero-better-bibtex/citing/

[2] Pandoc, User's Guide: Citations. https://pandoc.org/MANUAL.html#citations

[3] Espanso, Documentation: Basics. https://espanso.org/docs/get-started/

[4] NVIDIA, Parakeet TDT 0.6B v2 Model Card. https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2

[5] Hugging Face, Open ASR Leaderboard. https://huggingface.co/spaces/hf-audio/open_asr_leaderboard

[6] OpenAI, Whisper large-v3 Model Card. https://huggingface.co/openai/whisper-large-v3

[7] National Institutes of Health, The Use of Generative Artificial Intelligence Technologies is Prohibited for the NIH Peer Review Process. https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html

[8] Committee on Publication Ethics, COPE Position Statement: Artificial Intelligence and Decision Making. https://publicationethics.org/guidance/cope-position/artificial-intelligence-ai-decision-making