Alexandria Audiobook
A local AI audiobook generator uses Qwen3-TTS for multiple voices, connects to local or cloud LLMs, and exports MP3 or chaptered M4B.
Alternative to:
Alexandria Audiobook turns books into audio with separate narrator and character voices. It's for people who want to choose the cast and correct individual passages before exporting an audiobook.
- Script annotation: A separate LLM identifies speakers and delivery directions, with an optional review pass to correct annotation errors.
- Character voices: Use presets, clone a voice from a short recording, or design one from a written description; automatic persona generation can create and assign voices from the script.
- Voice training: LoRA training creates reusable voice identities that follow delivery instructions, and a dataset editor helps prepare text, emotion labels and audio samples.
- Passage editing: The browser editor lets you change speakers, text and delivery directions, preview audio, and regenerate selected passages without rendering the whole book again.
- Audiobook exports: Save a combined MP3, chaptered M4B, individual voice lines, or separate speaker tracks and labels for Audacity.
The built-in Qwen3-TTS engine runs locally. Script annotation requires an OpenAI-compatible API, such as LM Studio, Ollama or OpenAI. Local backends keep annotation and speech processing on your hardware; cloud annotation sends book text to the chosen provider. Speech can optionally use a remote Qwen3-TTS server. Model downloads require internet access.
The app is open source under the MIT license and runs on Windows, Linux and macOS. NVIDIA GPUs work on Windows and Linux; AMD GPUs work on Linux. Macs, including Apple Silicon, and AMD hardware on Windows use significantly slower CPU processing. GPU operation requires at least 8 GB VRAM. The app needs at least 8 GB RAM and recommends 16 GB. Docker supports NVIDIA server deployments.