Firebrat turns a technical PDF — the kind with circuit diagrams, register tables, and equations — into a narrated audiobook that keeps the right figure or formula visible and highlighted exactly while it's being talked about. Nothing gets flattened into noise. Novels can now get a different voice for every character (preview), and a free community library of copyright-free audiobooks is opening.
Most audiobook tools treat a PDF as a wall of text to flatten into speech. That breaks the moment a book says "as shown in Figure 6.65" or writes out an equation.
An LLM recognizes the mathematical relationship in the page text, writes real narration for it, and Firebrat renders and highlights the actual formula on screen at the exact moment it's spoken.
The diagram currently being described gets emphasized — a glow, a border — but every other figure and table for that section stays in view and tappable. Nothing is ever hidden from you.
Tap any sentence to jump straight there. Skip forward, skip back, jump sections — always, regardless of whether autoplay is on. Autoplay is a convenience layered on top, never a constraint.
Download or import a book once and the reader never needs a network again. Books are ordinary portable .tar.gz files — copy one to a phone over USB and open it directly, no server required at all.
Adjustable text scale, high-contrast highlighting, and short, sentence-length segments are core design decisions from day one — not accessibility bolted on after the fact.
Narration is synthesized with an open-source TTS model running on your own hardware — no per-minute cloud TTS bill, no audio ever leaving the machine that converts the book.
A free library of copyright-free audiobooks, hosted on the Internet Archive. Listen on the web or in the Android app, download for offline, rate and comment, or submit your own. The library is new and the first books are still being produced.
For novels, an LLM casts the characters and a speech model gives each one a distinct voice and emotion. The result is the standard package, so every client plays it unchanged. Preview: not yet verified on real books.
Opt-in accounts (admin, user, viewer), invites, model and run-pattern settings, and queue control from the web and Android apps. Implemented and unit-tested; not yet run on a live deployment.
Every section plays as a repeating cycle: the sentence sounds, the visual lights up. This is that loop, live.
The full technical breakdown lives on the how it works page — here's the shape of it.
Layout-aware parsing pulls out text, figures, and tables from the PDF, page by page.
An LLM groups pages into sections, writes clean narration, and identifies formulas worth showing.
Local TTS synthesizes each sentence, timed to the millisecond against the assembled audio track.
Everything is bundled into one .tar.gz — download it, import it, or share it directly.
Full technical textbooks — hundreds of pages each — converted end to end and served from the companion server.
"I'm dyslexic. Sustained silent reading of a 600-page technical textbook is genuinely exhausting and painful for me. Listening is how I actually get through dense material — but a technical book can't just be read aloud, the diagrams and formulas are load-bearing. Firebrat is the reader I built so I could stop choosing between the two."
— from the project READMEFirebrat is free, open source, and self-hosted. Convert a PDF on your own hardware, then read it on your phone — online or off.