Drop a PDF here or click to browse
Most papers convert in seconds with no download. Pages the extractor can’t linearize (dense stacked equations, scans) use the Nougat model — downloaded once, then cached; a browser with WebGPU (Chrome or Edge) makes that part much faster.
How it works
- Extract — the words are read directly from the PDF’s own text layer (pdf.js), so the prose is exactly what the authors wrote — nothing is guessed or re-recognized.
- Rebuild the math — equations are reconstructed into LaTeX from the glyphs on the page: symbols by Unicode, variables by their math fonts, exponents and subscripts by their position and size.
- Fall back where needed — a page the extractor can’t linearize (heavily stacked equations, scanned pages) is handed to the Nougat neural model, running on your CPU/GPU. Only those pages take the slow path.
- Speak-ready — an open rule library turns LaTeX into plain English (so "x^2" is read "x squared" and "bab⁻¹" is read "b a b inverse"). Your feedback improves it.
- Read aloud — your browser speaks the paper while highlighting each word on the PDF itself. Click anywhere on the page to start listening from there. Citations, footnotes, and the bibliography are kept out of your ears.
🔒 Nothing is uploaded. The conversion happens entirely on your device.