Home

Alexandria Audiobook2 Wiki

Alexandria Audiobook2 turns a book into an audiobook with a different voice for every character. A language model works out who says each line and how it should sound; a text-to-speech model (Qwen3-TTS) reads it; and you get an MP3 or a chaptered M4B.

This fork measures every choice it makes. The research pages below say what was measured and how; the user-guide pages say how to use the app.

Start here

• Use the app: Install, connect an LLM, then follow Your first audiobook. Start with the original short sample, assign CustomVoice presets, render, merge and download an MP3 before trying a full book.

• 中文入门：安装与第一本有声书。本 Wiki 的详细页面目前为英文。

• Research or develop: the research pages below, GOALS, RECIPES and contributor instructions preserve the measurements, reproduction steps and development workflow. None is a prerequisite for making a book.

• 日本語：日本語 Wikiとインストール・最初のオーディオブック。

Using the app

• Setup and serving — connecting a language model, and serving settings that were measured to work.
• Script generation — the three passes that turn text into a script, the review pass, and prompts.
• Voice types — CustomVoice, Clone, LoRA and Voice Design, and when to use each.
• Voice reference — a vocabulary for voice descriptions and delivery instructions, with test results.
• Batch generation — how rendering is batched, and how to tune it.
• Editor & export — editing lines, render modes, merging and export.
• Troubleshooting — common problems and fixes.

Building voices

• Dataset builder — build a voice-training dataset with a preview of every sample.
• Training guide — train a LoRA voice, with settings that worked.

Reference

• API reference — HTTP examples in curl, Python and JavaScript.

Research

• Which model for your card — the best language model for each amount of GPU memory, with speeds.
• Results — model, quant, prompt and adapter results in one place.
• Prompts and adapters — the rules adapters must follow, and when an adapter helps.
• Evaluation recipes — how an adapter evaluation is run.
• Thunder operations — running jobs on remote GPUs.
• Hugging Face releases — what a public release must include.

The one rule for adapters

An adapter must be evaluated with the prompt it was trained on. A result under a different prompt is a transfer test, not a result for that adapter.

Checking an installation

The repository includes a live API test suite. The safe way to run it starts a throwaway copy of the app with its own empty data folder, so your books and voices are never touched:

cd app
python run_isolated_api_tests.py          # quick tests — no TTS or LLM needed
python run_isolated_api_tests.py --full   # also tests generation — needs a GPU and an LLM

To test an app that is already running, use python -m tests.test_api --url http://127.0.0.1:<port> from the app folder — but note that it creates and deletes uploads, scripts and voice settings in that app's data. (Running python tests/test_api.py directly fails with ModuleNotFoundError: utils; it must be run with -m from app.)

• Quick mode reports its own run/skip counts and covers settings round-trips, upload, the script library, voice settings, chunks, status polling, voice design, LoRA listings, the dataset builder and error handling, without loading any model.
• Full mode adds script generation, audio generation, batch rendering, voice design previews and LoRA testing.

Passing quick mode checks the model-free API paths exercised by that suite. It does not rule out application bugs in generation, model loading, GPU execution or provider integration. Use the failing task's logs to distinguish those causes; full mode covers additional paths but is not an exhaustive test of every model or book.