Analysis and the Analyzer
This is where Castwright reads your book. Once a manuscript is saved, the analyzer works through it chapter by chapter — finding every character who speaks, and tagging each line with who's saying it — before a single second of audio gets rendered. Get this step right and everything downstream (casting, voices, the performance itself) inherits that accuracy; it's worth understanding what's actually happening on this screen rather than just watching a bar fill.
Under the hood it's a phased pipeline, not one giant pass: chapter boundaries are found first with a measured, observed-rate ETA rather than a guess; each chapter's cast is detected on its own, so one difficult chapter's failure doesn't torch the whole run; and every claimed line of evidence is checked back against your actual source text before it's trusted. The Analysing screen streams all of this live — phase, percentage, ETA, and the cast roster growing in front of you as it's discovered.
The model picker on the upload screen groups two kinds of analyzer:
- Local Ollama (on-device) — free, private, runs on your own GPU/CPU. Whatever tags you've pulled with Ollama show up here automatically.
- Gemini / Gemma (cloud) — Google's free-tier API. Useful on a low-VRAM machine, or when you'd rather not tie up local compute.
See Installing Castwright for setting up either path (pulling an Ollama model vs. adding a Gemini API key).
A cloud analyzer also unlocks a pipelined two-model split on the Analysing screen itself — one model races ahead detecting characters (Phase 0) while a second, more careful model trails behind verifying attribution (Phase 1), gated by a warm-up lag so the second model always has a roster to work against before it starts. Configure each phase independently, or leave both on "(use server default)" for the ordinary single-model path.