From E-Book to Full-Cast Audiobook on One GPU

What it does

Alexandria is a browser-based pipeline that ingests a book or script, uses an external LLM to annotate speakers and emotional direction, then renders the whole thing into audio via a built-in Qwen3-TTS engine. It outputs finished MP3s, chaptered M4Bs, or per-speaker WAV tracks you can drop straight into Audacity. The web UI lets you edit every line and re-render individual chunks without starting over.

The interesting bit

The project treats the LLM as a casting director and script supervisor: it parses narration and dialogue into JSON, generates voice descriptions for each character, and can automatically assign cloned or designed voices. It even passes character context between chunks to keep names and styles consistent across long books. That’s the kind of glue code that sounds boring until you realize it’s doing the work of a small audio production team.

Key highlights

• Built-in Qwen3-TTS runs locally; no external TTS server needed (though you must bring your own LLM API)
• Voice cloning from 5–15 second samples, text-to-voice design, and LoRA training for persistent character voices
• Exports to MP3, chaptered M4B (Audiobookshelf/Apple Books compatible), or multi-track Audacity projects with per-speaker WAVs and labels
• Smart chunking groups consecutive lines by speaker up to 500 chars for natural flow
• Optional torch.compile codec optimization and batch generation for 3–6x real-time throughput

Caveats

• Requires 8 GB VRAM minimum (16 GB+ recommended); AMD and Apple Silicon are limited to CPU on Windows/macOS respectively
• TTS model weights (~3.5 GB each) download on first use, and the initial batch has a 30–60 second warmup
• The app does not include an LLM; you must run LM Studio, Ollama, or an OpenAI-compatible server separately

Verdict

Ideal for hobbyist authors and tinkerers who want a fully voiced audiobook without hiring narrators, provided they have the GPU patience. If you’re looking for a one-click mobile app or lack a discrete GPU, this is not your tool.