Engines & models
Draft — verify against the current release.
Castwright ships with three synthesis engines. You do not need all three — Kokoro alone is enough to produce a full-cast audiobook.
Kokoro (default)
Kokoro is the default engine and the one Castwright uses for generation unless you choose otherwise. It is always resident in memory once the weights are installed, so there is no wait for a model to load between chapters.
- Languages: English only (28 voices).
- Voice names: prefixed
af_,am_,bf_,bm_(American/British female/male). - VRAM: ~1 GB while generating.
- When to use it: for every character when you want fast, reliable English synthesis.
Qwen (voice design)
The Qwen engine is used for designing custom voices. When you open the “Design voice” panel for a character, Castwright loads the Qwen VoiceDesign model, generates a voice from your description, and lets you compare the result with the default. You can accept the design, try another description, or discard it.
Qwen is also the synthesis engine for any character whose voice was designed with it. It is loaded on demand, not at startup.
- Languages: multilingual.
- VRAM: ~3.7–5 GB while generating, depending on the model size; the design model unloads automatically once you leave the cast-review screen.
- When to use it: when you want a character voice that is precisely tuned — a specific accent, age, affect, or timbre that no preset covers.
Coqui XTTS v2 (optional)
Coqui XTTS is an optional multilingual engine. You load it explicitly via the Model Manager; it is not installed by default. It currently performs with pre-designed voices; once voice cloning ships, it will also carry cloned voices.
- Languages: multilingual.
- VRAM: ~4 GB while generating; loading Coqui unloads the Ollama analysis model to reclaim memory.
- When to use it: when you want multilingual pre-designed voices today. Narrating in your own voice — cloning from a short sample — is in development and ships in an upcoming release.
The analysis model
Before any voices are generated, Castwright reads the manuscript to work out who says what. That analysis runs on a model on your own machine — a local Ollama model — or, if you prefer, the free Gemini tier. You choose which installed model does the reading, and you can pull more from the Model Manager; the analysing screen names the model actually in use, so you always know what is reading your book. Because the analysis runs separately from speech, you can pair a quicker model with a more careful one across the passes without tying up the GPU your voices need.
The Model Manager
The Model Manager lives under the account menu (or at #/models). It shows all installed and available engines, lets you install or remove voice weights, and displays current GPU memory usage. Each engine shows its honest state — package, weights, and integrity — and one that needs putting right reads Needs repair, with a one-click Repair that reinstalls only what is missing and restarts it. If an engine fails to load, the error appears here first.
VRAM at a glance
These are run sizes — the VRAM in use while generating, not the download size:
| Engine | VRAM while generating |
|---|---|
| Kokoro | ~1 GB |
| Qwen — Fast (0.6B) | ~3.7 GB |
| Qwen — Higher quality (1.7B) | ~4.7–5 GB |
| Qwen — Voice Design | ~3.7–4 GB (base auto-unloaded) |
| Coqui XTTS v2 | ~4 GB |
Two more expressive engines are planned for bigger cards — IndexTTS-2 (per-line emotion, likely a 12 GB card) and Fish Audio S2-Pro (16 GB card tier) — landing as bigger-card support matures. See system requirements for how these fit alongside the analysis model on 6, 8, and 12–16 GB cards.