Your e-reader finally learned to talk

What it does Abogen is a desktop and web tool that ingests EPUBs, PDFs, plain text, markdown, or subtitle files and spits out narrated audio with matching captions. It targets content creators making voiceovers and readers who want audiobooks without the Audible markup. Under the hood it uses Kokoro-82M, a compact neural TTS model, and offers GPU acceleration on NVIDIA and AMD hardware.

The interesting bit The subtitle synchronization is the quietly hard part. Abogen generates timed captions at variable granularity — by sentence, by comma clause, or even word-by-word — which matters if you’re feeding the output into video editors or accessibility tools. There’s also a voice mixer for blending speaker profiles, and a batch queue for processing multiple books overnight.

Key highlights

• Dual interfaces: a PyQt6 desktop app (abogen) and a Flask web UI (abogen-web) with a feature gap — the web version currently gets LLM text normalization, Supertonic TTS, and Audiobookshelf integration first
• Output formats include M4B with chapter markers, plus WAV, FLAC, MP3, and OPUS
• Chapter-aware processing: extract specific chapters from EPUBs or page ranges from PDFs, save individually or merged
• Cross-platform with Windows, macOS (Intel and Apple Silicon), and Linux support, though AMD GPU acceleration requires Linux due to ROCm limitations on Windows
• ~3,000 characters processed in ~11 seconds on an RTX 2060 Mobile, per the author’s demo

Caveats

• The web UI is ahead of the desktop app in features, and the README notes the gap is being backfilled
• macOS and Windows installs require espeak-ng as a separate system dependency, and the Windows path involves either a bundled installer script or wrestling with CUDA-specific PyTorch versions
• The “powerful” descriptor in the README is doing heavy lifting; this is essentially a polished wrapper around Kokoro-82M with file-format plumbing

Verdict Worth a look if you produce captioned video content or want private, offline audiobook generation. Skip it if you need production-grade emotional range in narration — Kokoro-82M is competent but not dramatic.