A multi-voice AI audiobook generator, Alexandria transforms books into audiobooks using LLM-driven script annotation and Qwen3-TTS. It targets users needing automated, high-fidelity audiobook production with advanced voice customization, offering unique character voices, cloning, and fine-tuning capabilities.

How It Works

Alexandria employs an AI pipeline that first uses an LLM to parse book text into a structured JSON format, identifying speakers, dialogue, and TTS instructions. An optional second LLM pass refines annotations. The core is the Qwen3-TTS engine, which can run locally or remotely, synthesizing speech with per-line style control. Novelty lies in its comprehensive voice generation suite: cloning from short audio samples, designing voices from text descriptions, and persistent voice identity training via LoRA fine-tuning.

Quick Start & Requirements

Installation is recommended via Pinokio, or a Google Colab notebook is available. A separate OpenAI-compatible LLM server (e.g., LM Studio, Ollama) must be running. A GPU with 8 GB VRAM minimum (16 GB+ recommended) is advised for optimal performance, supporting NVIDIA (CUDA 12.8+) and AMD (ROCm 6.3+ on Linux). CPU mode is available but significantly slower. Requires ~20 GB disk space and 16 GB RAM recommended. Initial TTS model downloads (~3.5 GB per variant) occur on first use.

Highlighted Details

• Advanced Voice Synthesis: Features voice cloning from 5-15 second references, text-based voice design, and LoRA fine-tuning for custom, persistent voice identities.
• LLM-Powered Scripting: Automates script annotation, including character dialogue, narration, and context-aware non-verbal sounds, with an optional review pass for error correction.
• Flexible Export: Generates a single MP3 audiobook or prepares multi-track Audacity projects with per-speaker WAV files and labels for detailed editing.
• Performance: Achieves 3-6x real-time throughput using batch rendering with codec compilation enabled.

Maintenance & Community

The project notes a recent surge in user attention, which may lead to slower issue response times. No specific community channels (Discord/Slack) or prominent contributors are detailed in the README.

Licensing & Compatibility

Licensed under MIT, Alexandria is generally compatible with commercial use and closed-source linking without significant restrictions.

Limitations & Caveats

A separate LLM server is a mandatory prerequisite. AMD GPUs on Windows and Apple Silicon Macs are limited to CPU processing, resulting in substantially slower performance. Initial TTS model downloads require stable internet connectivity.