Firebrat

Most audiobook tools treat a PDF as a wall of text to flatten into speech. That works for a novel and falls apart the moment a book says "as shown in Figure 6.65" or writes out an equation — the audio either skips the thing you actually need to see, or reads a LaTeX expression aloud as noise.

Firebrat is built for readers who need the audio to be the primary way through a book, not a background convenience: dyslexic readers, people with low vision or ADHD, or anyone who processes spoken language better than dense printed text. Accessibility — adjustable text size, high-contrast highlighting, short-segment pacing — is a core design decision, not an add-on bolted on afterward.

The pipeline converts a PDF into a self-contained book package: extraction (marker-pdf + PyMuPDF), LLM compilation that groups pages into sections, writes narration text, and authors LaTeX for the formulas it recognizes, then per-segment TTS with sample-accurate timestamps (Chatterbox).

The companion Flutter app reads a section at a time, keeps every figure and table for that section visible in a scrollable gallery, renders and highlights the formula currently being spoken, and lets you jump around freely — by tapping any segment, prev/next controls, or autoplay. A FastAPI server distributes packages; download once, then fully offline. Docker images (CPU / GPU / API-only) are published on Docker Hub.