Research papers are a specific kind of hard to read: dense, citation-heavy, full of field-specific jargon, and often formatted in two columns with figures interrupting the flow. Listening does not fix all of that, papers are still papers, but it changes where and how you can get through a first pass, and that is often the actual bottleneck when you have a stack of papers to get through before a meeting or a deadline.
Getting a paper in: arXiv links and PDFs
The simplest path for a preprint is pasting the arXiv link straight into ReadLoud as a new document, the same way you would paste any other link. ReadLoud extracts the page content, similar to a reader view. If you already have the PDF downloaded, whether from arXiv, a journal, or your institution's library access, you can upload it directly instead (up to 40 MB and 900,000 characters, which covers essentially any paper).
One real limitation to know upfront: extraction reads a document's actual text layer. A paper that only exists as a scanned image, with no underlying selectable text, has nothing for the extractor to read, so it will not produce usable audio. Most modern PDFs, from arXiv, journal publishers, and preprint servers, have a proper text layer, so this mostly affects older scanned material.
Citations get out of the way
Academic writing is full of citation markers, [12], (Chen et al., 2021), sometimes stacked three or four deep in a single sentence. Read literally out loud, these are genuinely disruptive, a voice reading "bracket twelve" mid-sentence breaks your attention every time. ReadLoud drops citation brackets and reference-style markers during extraction so the sentence reads as the author intended, as prose, not as prose interrupted by footnote numbers. Headers, footers, and page numbers get the same treatment. If you want to know exactly what was dropped, skipped text is shown dimmed on the page rather than silently vanishing, so nothing disappears without a trace.
Jargon and the pronunciation dictionary
Papers are full of field-specific terms, author surnames, model names, and acronyms that a general-purpose voice will sometimes mangle. ReadLoud has a pronunciation dictionary for exactly this: if a voice keeps getting a term wrong, you can correct it once, and that correction applies consistently going forward rather than needing to be fixed every time the term appears. This is a small feature but it matters more for academic material than almost anything else ReadLoud handles, since papers repeat specialized vocabulary constantly.
The equations caveat, stated honestly
This is worth being direct about rather than glossing over: equations do not translate to speech in any useful way. A voice cannot narrate an integral, a matrix, or a multi-term derivation the way a professor would when talking through it on a whiteboard. Depending on how an equation is encoded in the PDF, expect it to be skipped, rendered as scattered symbols, or read as a string of characters that does not mean anything spoken aloud. This is not a gap unique to ReadLoud, it is a general limitation of turning mathematical notation into audio at all.
The practical approach: use listening for the prose, the introduction, related work, methodology description, discussion, and conclusion, which is usually the majority of a paper's words, and switch to looking at the actual page for equations, tables, and figures. Papers are not built to be consumed purely by ear, and treating listening as a first-pass or a way to cover the connective narrative, rather than a full substitute for reading the technical core, is the realistic way to use it.
Voices and speed for dense material
For papers, a calm or briefing-style delivery on one of the Gemini-based voices tends to suit the material better than something upbeat. You can preview voices before committing, and switch mid-document without losing your place, useful if a paper's tone shifts between a dense methods section and a more readable discussion.
Speed matters differently here than for casual reading. Dense academic prose often benefits from a slower pace, or at least a pace under whatever you would use for a news article, since the information density per sentence is higher. ReadLoud runs from 0.5x to 4x with pitch preserved, so slowing down does not distort the voice, and you can adjust speed on the fly as a paper gets denser or eases up.
Word highlighting for cross-referencing
ReadLoud's highlighting is synced to the real audio through forced alignment, with a sentence tint and a click-to-jump feature on any word. For papers specifically, this is useful when you want to glance back at a sentence you just heard to check a specific term or number rather than relying purely on memory. Focus mode dims everything but the current sentence, which helps on a page that also has figures and captions competing for attention.
Your library, across a stack of papers
Reading for a literature review or a qualifying exam usually means working through many papers over days or weeks, not one sitting. ReadLoud keeps a library of every document you add, with progress tracked and resumed at the exact sentence you stopped at, synced across devices. That means a paper you started on your laptop at your desk can be picked back up on your phone during a commute, at the exact point you left it, rather than needing to relocate your place by scrubbing through the audio.
You can also switch voices mid-document without losing your place, which is useful across a stack of papers from different fields or writing styles, a dense statistics paper and a more narrative case study do not need to sound the same.
Trying it on a paper you already have
ReadLoud gives 1 minute of listening with no account, and 20 minutes with a free account using email sign-in and no password. That is enough to run one real paper through the whole flow, pasting the arXiv link or uploading the PDF, picking a voice and delivery style, and seeing how the citation skipping and pronunciation dictionary actually behave on your own material, rather than judging from a generic example.
Where this fits into an actual reading workflow
Realistically, listening works best as triage and first-pass reading: getting through the introduction and related work for a stack of papers before deciding which ones need a close, on-page read of the technical sections. It is not a replacement for sitting with the math. Used that way, it saves real time without pretending papers are simpler than they are.
For more on academic reading specifically, see the research papers page.