Dictation for Academics, Researchers and Theses

Dictation is a good tool for the drafting stage of academic work and a poor one for citations, equations and layout. Speak the argument, the methods and the discussion in full sentences, then leave the references, the math and the tables to the keyboard or to tools built for them. A dictated draft also reads more like speech than writing, so it needs a second, deliberate editing pass before it is ready for a journal, a committee or a funder.

This guide covers what researchers dictate and what they type, how to build an argument by speaking it, how to handle citations with a reference manager, what the current tools can do with math and LaTeX, how to use dictation for field and lab notes, and which confidentiality rules apply to unpublished work and peer review. Our overview of dictation for work and study places it among the other guides in this series.

Dictation fits the drafting stage of academic writing, not the citation and formatting stage

Speech produces words much faster than typing. In a controlled study on an iPhone, English text entered by speech reached 153 words per minute, against 52 words per minute on the built-in keyboard ^{1 2}. The study used short phrases in a quiet room, so it measures the best case for speech, not a thesis chapter written across a noisy office.

Speed is only part of the picture. The time you save on composing sentences can disappear in correction. In a study of dictation with a large-vocabulary recognizer, Vertanen reports that correcting errors dominates the time spent on the task ³. Plan for a dictated chapter to take a full editing pass, not a quick read.

Dictation is also quick to learn, though the evidence comes from older studies. In a 1978 study published in Science, novice dictators learned within a few hours to dictate with the speed and quality with which they wrote letters ⁴. Letters are not research papers, so treat the result as a sign that the skill is quick to learn, not as proof for a thesis.

The table below sorts the common parts of a research document by the tool that suits them best. The advice is ours, and the reasons appear in the sections that follow.

Our own app, Yap, which we make, is an on-device dictation tool for Mac and Windows. It is built for drafting prose, and it does not replace the reference manager, the equation editor or the layout tools above. Where an institution asks you to use a specific writing environment, follow its rules first.

Can you dictate a thesis or dissertation?

Yes, in the sense that nothing in dictation software stops you from speaking chapter drafts. A thesis is mostly prose, and prose is what dictation does best. Our question on dictating a thesis goes through the practical steps.

Two practical limits matter more than the software. First, Microsoft says the dictation feature in Word is only available to Microsoft 365 subscribers ⁵, and Microsoft's description of dictation in Microsoft 365 calls for a microphone and a reliable internet connection ⁶. Second, if dictation is part of your accommodation, the rules come from your institution. Under Section 504, a college that receives federal money must make the modifications to its academic requirements that are necessary so they do not discriminate against students with disabilities ⁷, and your disability office decides whether a dictation accommodation covers extended, unsupervised writing such as a thesis. Ask them before you rely on it, and keep the approval in writing.

Talking through an argument before writing it produces a clearer first draft

Much of the work of writing goes on deciding what to say, not on typing it. In the same 1978 study, about two-thirds of the time spent composing a letter was planning, whether the letter was written or dictated ⁸. Speaking your argument aloud is a way of doing that planning in a form that is quick to produce.

Outlining also pays. A 1988 study found that preparing a written outline before writing led to higher-quality documents than not outlining, as rated by judges ⁹. Kellogg's 2001 experiments found that planning, translating ideas into sentences and reviewing compete for one general-purpose working memory resource ¹⁰. In our reading, dictation takes the keystrokes out of translation, but it does not remove the need for planning and review. Plan the argument in a short spoken outline, then dictate each section against that outline.

A practical sequence for a paper or chapter:

1. Record a spoken outline of the claim, the three or four reasons for it, and the evidence for each reason.
2. Dictate one section at a time, from that outline, in complete sentences, and mark every place where a citation or number belongs with a visible placeholder.
3. Rewrite for academic register before you send the draft to anyone.
4. Insert citations and check every number against its source.

Step 3 matters because speech and writing differ in how many content words they pack in. Eggins, as cited by Ishikawa in 2015, found in a corpus analysis that lexical density (the share of content words, such as nouns, verbs and adjectives, among all words) was 33 percent in speech and 42 percent in writing ¹¹. Ure, also cited by Ishikawa, put the lexical density of speech below 40 percent and that of writing above it ¹². A dictated draft may sit on the spoken side of that line, so measure its density against the writing you want, and revise the sentences that read as speech.

The speed gain is real, and so is the need for revision. Snider's 2002 dissertation studied postsecondary students with learning disabilities in written expression. They wrote higher-quality essays, at a faster rate, with speech recognition than by keyboarding ¹³. Essay length did not differ significantly among the three ways of writing in that study ¹⁴. The study covers students with learning disabilities only and dates from 2002, so its recognition software is long superseded. It supports the claim that dictation can help composition, not that it will improve every academic writer's output.

Citations belong in the reference manager, inserted by keyboard or voice command, not spoken in full

Do not dictate a citation in full. Author names, years, page ranges and DOIs are exactly the kind of word that recognizers get wrong. In a 2024 study, Apple researchers reported that end-to-end speech recognition still has a significant error rate for entity names that appear infrequently in its training data ¹⁵. Author names are often rare words of exactly this kind.

The pattern that works is to dictate the sentence, leave a placeholder where the citation belongs, and insert the citation from your reference manager. Zotero is a reference manager with a word processor plugin. Its official documentation says that in most word processors it is possible to assign keyboard shortcuts to the plugin's commands, such as add citation and edit citation ¹⁶. The steps depend on the program, so follow Zotero's documentation for your word processor and version.

After every dictated section, check the reference list against the text. Read the author names and years aloud from the manager, not from the transcript, because a misheard name is easy to miss on the screen.

Math and LaTeX can be dictated, but only with specialist tools

Standard dictation can insert symbols, not the structure of an equation. Microsoft's English command list for Dictate in Word includes a Mathematics table, with phrases such as plus sign, minus sign, multiplication sign, division sign and equal sign ¹⁷. Apple's Mac dictation commands also let you enter symbols for math and currency ¹⁸.

The command lists we read name symbols, not layout such as fractions, superscripts, matrices or numbered equations. Check the complete list for your app before you rely on it, and assume that anything with structure needs the keyboard or an equation editor.

Specialist work on speech to LaTeX is more advanced. A 2025 preprint, Speech-to-LaTeX, released a dataset of annotated audio of equations and sentences, and it reports results on the MathSpeech benchmark and on a benchmark of its own. On its own benchmark of spoken equations, its models outperform the MathSpeech model by more than 36 percentage points ¹⁹. On the existing MathSpeech benchmark, its models reach a character error rate of 28 percent, against 30 percent for MathSpeech, which the authors describe as comparable ²⁰. A character error rate counts wrong, missing and extra characters, so a lower figure is better. The 36-point result uses the authors' own test set, and the preprint has not been peer reviewed.

For day-to-day work, the practical options are:

• Dictate the sentence, insert the equation. Say "the loss is the sum over all samples", then add the equation in your LaTeX editor or equation editor.
• Try a community command set for Talon Voice. Community projects for math and LaTeX have been built for Talon, such as mathfly. We did not test them, so treat their coverage as unknown until you try them on your own equations.
• Use a research system only for trial runs. The 2025 preprint is evidence of progress, not a product you can rely on for a submission.

Whatever the method, read each equation against its source line. A recognizer that is nearly right on a symbol can still change the meaning of the whole line.

Dictated field and lab notes capture observations before memory fades

The evidence for dictated field notes is thinner than the evidence for drafting. We found no peer-reviewed methods study of voice field notes in our checks, so the advice below rests on what dictation and transcription research says about accuracy, correction and privacy. Our question on dictating field notes gives the step-by-step workflow.

The case for speaking the notes straight after a session is about memory, and that case is a reasonable one. The case for transcribing them with software needs checking. Listening back to recognized text is not a reliable check: in an in-lab study, participants missed identifying over 50 percent of speech recognition errors when they listened to the audio output of the recognized text ²¹. Any transcript of field notes should be checked against the recording, not against the transcript itself.

Transcription tools can also invent text. A 2024 study of Whisper, an open speech recognition model, found that roughly 1 percent of the audio transcriptions in its dataset contained entire hallucinated phrases or sentences that did not exist in the audio ²². An invented sentence in a field note can look as plausible as a real one, so check any passage that matters to your analysis.

Error rates also differ between speakers. Across five commercial speech recognition systems, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers ²³. The gap is a reason to plan extra checking for interviews with speakers whose dialect or speech patterns differ from the training data, and to report the method in your ethics or methods section. The study measures these errors as word error rate, the standard measure defined in our glossary.

Privacy comes before accuracy with participant recordings. On Apple devices, Apple says dictation sent to its servers is not stored unless you opt in to Improve Siri and Dictation ²⁴. Apple also says only its own employees, subject to strict confidentiality obligations, are able to access audio interactions with Siri and Dictation ²⁵. Those are Apple's rules for Apple's services, not a substitute for your ethics approval. Your consent forms and data management plan decide where participant audio may go.

A practical workflow for field notes:

1. Record a voice note straight after each session, while the details are fresh.
2. Transcribe it only with a tool your ethics approval allows, and keep participant audio on a device or service the approval names.
3. Check every quoted passage and every name against the audio.
4. Mark anything the transcript invented, and remove it before the notes enter your analysis.

Unpublished research and peer-review material limit which cloud tools you may use

The rules for unpublished work depend on who owns the text. For peer review, the rules are clear and strict. NIH revised its confidentiality agreements for peer reviewers in 2023 so that reviewers are prohibited from using AI tools in analyzing and critiquing NIH grant applications and R&D contract proposals ²⁶. The rule governs reviewers, not applicants, and a summary of the notice quotes it rather than the NIH page itself, so check the notice before you rely on the wording.

For your own unpublished drafts, the question is which service handles the audio and the text. Apple says your device indicates, in Keyboard settings, whether dictated audio and transcripts are processed on the device or sent to Apple's servers ²⁷. Check that setting before you dictate unpublished results, and treat any service that sends audio elsewhere as a question for your data management plan.

Three checks before you dictate material that is not yet published:

• Where the audio goes. Use on-device processing where you can, and check the setting rather than assuming it.
• What your funder and journal say. Peer-review rules, such as NIH's for reviewers, can limit what you may do with the proposals you review, so check them before you read or summarize one with a tool.
• What your ethics approval allows. Participant data is governed by the approval, which may name the tools you can use.

Our overview of voice privacy explains how these settings differ between services.

Researchers writing in a second language can draft by voice and edit with care

We found no study in our checks that measures dictation quality for academic writers working in a second language, so we cannot give a rate for them. The safe approach is to treat the draft as a spoken draft, check it more carefully, and use the tools' own settings to reduce errors.

We have no error rates for accented second-language academic speech. The speaker-group gap above is a warning sign rather than a measure, and the same risk may apply when the accent and phrasing sit outside the training data. Two habits help. Add your field's terms and names to your tool's custom vocabulary or dictionary, if it has one, and read each dictated section aloud against the source before you submit it. Our question on dictation for non-native English speakers at work covers the workplace version of the same steps.

Lexical density is a measure of how compact the prose is, and speech is less dense than writing ¹¹. A dictated draft may therefore need tightening into denser phrasing, whatever your first language. Reading the draft back and revising it into that phrasing is the step those figures point to.

A workable split: dictate the draft and type the apparatus

Use dictation for the parts of academic work where speech is fast and the argument matters most. Use the keyboard and specialist tools for the parts where a small error changes the record. For most researchers, the sequence is:

1. Talk through the argument as a spoken outline.
2. Dictate each section in complete sentences, with placeholders for citations, equations and figures.
3. Rewrite the draft for academic register, because speech is less dense than writing.
4. Insert citations from your reference manager and equations from your editor.
5. Check names, numbers and equations against their sources, and check every transcript of an interview or field note against the recording.

The same split applies whether you are writing a journal article, a grant or a thesis chapter. Our guides on writing a book by dictation, speech to text for students, dictation productivity at work and dictation for journalists and content creators cover the same drafting method for other kinds of writing. For a broader view of how speech recognition works and where it fails, see our overview of transcription and our overview of dictation for work and study.

If you want a free tool to try first, online dictation runs in your browser, and audio to text can transcribe a recorded interview for checking. Both need the same checks described above.