**Alexandria** turns a book or novel into a fully-voiced, multi-character audiobook, and the interesting part is how much of the work happens before any audio is synthesized. It is MIT-licensed, published by **Finrandojin**, and has reached **972 stars and 119 forks** since going public on **3 February 2026**. The README opens with a note that the project has seen a sudden surge of new users and that the maintainer may not answer every issue promptly — an unusually candid piece of expectation-setting for a solo project at this scale. The pipeline starts with an **LLM, not a TTS engine**. Any OpenAI-compatible endpoint works — LM Studio, Ollama, OpenAI itself — and the model parses raw text into JSON annotated with speakers, dialogue, and per-line *instruct* directions describing how a line should be delivered. An optional second LLM pass reviews that output and fixes the errors this step reliably produces: attribution tags left inside dialogue, narration misattributed to a character, over-split narrator entries, invalid instruct fields. **Persona Generation** then has the LLM analyze the finished script, write a voice description for each character, generate reference audio through the Voice Designer, and assign clone voices automatically — one click from raw script to a cast. **Speaker Aliases** map name variants ("YOUNG ELENA" → "ELENA") onto a single voice, and context is preserved between chunks by passing the character roster and the last three script entries forward, which is what keeps names and style from drifting across a long book. Voice generation runs on a **built-in Qwen3-TTS engine** with no external server required, though a remote Qwen3-TTS Gradio server can be used instead. There are 9 pre-trained voices with instruct-based emotion control, zero-shot **voice cloning from a 5–15 second reference clip**, a **Voice Designer** that builds a voice from a text description, and **LoRA fine-tuning** with an interactive dataset builder for creating persistent voice identities. Pre-trained LoRA adapters ship out of the box. Ten languages are supported plus auto-detect, and the LLM writes non-verbal vocalizations ("Ahh!", "Mmm...") with context-aware directions, with configurable pauses between speakers (500 ms default) and within a speaker's own run (250 ms). Throughput is the practical constraint, and the README publishes a table rather than a slogan: sequential mode runs at roughly **1× real-time**, batch mode without codec compilation about **2×**, and batch mode with `torch.compile` codec compilation **3–6×** on batches of 20–60 chunks. Export options are genuinely thorough — a combined MP3, individual MP3s per line for DAW work, a one-click **Audacity** zip with per-speaker WAV tracks plus an LOF project file and labels, and a chaptered **M4B** for Audiobookshelf, Apple Books, or VLC. A browser-based five-step editor allows editing speaker, text, and instruct on any line and regenerating individual chunks without re-rendering the whole book. The hardware story is where prospective users should look carefully. Each TTS model occupies about **3.4 GB of VRAM**, with 8 GB the stated minimum and 16 GB+ recommended; remaining VRAM determines batch size. NVIDIA on Windows and Linux gets full support (driver 550+, CUDA 12.8, flash attention included), and AMD on Linux works through ROCm 6.3+. But **AMD on Windows, Apple Silicon, and Intel Macs are CPU-only** — MPS acceleration is explicitly not supported, described as functional but slow. Given the 3–6× figure applies only to GPU batch mode, Mac users should read "functional" carefully before committing to a full-length novel. Installation is via Pinokio, Docker, or a Colab notebook on a free T4 for anyone without suitable local hardware.