System requirements
Will it run on your machine?
A gaming PC or gaming laptop with a 6 GB GPU gets you started; 8 GB is the sweet spot. Castwright runs your audiobooks locally — the GPU does the heavy lifting. Here's exactly what runs on what.
Minimum
- 6 GB GPU — runs the default Kokoro (English) path with cloud or a small local analyzer; heavier designed-voice work is tight but works. NVIDIA benchmarked (RTX 3060 class or newer); Apple silicon supported
- Windows 10/11 or Linux with an NVIDIA GPU; macOS 12+ on Apple silicon (runs on the Mac's GPU — slower than the NVIDIA path)
- AMD GPU — experimental preview: Qwen and Coqui on ROCm, Kokoro on the CPU, with an automatic CPU fallback if the ROCm install doesn't take. Needs a ROCm-supported card and a recent driver
- Audio assembly — a free tool called ffmpeg, installed once with a single command (see the install guide)
- A few GB of disk for voice models + your library
Recommended
- 8 GB VRAM — the recommended path: designed voices, the Higher-quality tier, and comfortable headroom
- 12–16 GB VRAM — analysis and voices run at the same time (faster), and it's the tier the planned bigger-card engines target
- 16 GB RAM or more (the synthesis pipeline is memory-hungry)
- SSD storage — model loading and audio writes are faster on NVMe
What runs where
Character analysis (identifying who says what in your book) can run on a local Ollama model, or on the free Gemini API tier — so the GPU stays free for speech synthesis. CPU-only analysis is possible, though slower.
Speech generation (the voices themselves) runs on your GPU — NVIDIA is the benchmarked path; a 6 GB card runs the default voices, 8 GB is comfortable. Apple silicon runs via the Mac's GPU, slower. CPU-only generation works too, just slowly.
AMD (experimental) — Qwen and Coqui run on ROCm and Kokoro on the CPU (DirectML can't run the Kokoro model). If the ROCm install doesn't take, Castwright completes a working CPU setup and tells you it's on CPU. It's a preview; for full speed, an NVIDIA or Apple-silicon machine is the steady path.
| Engine | VRAM while generating | 6 GB | 8 GB | 12–16 GB |
|---|---|---|---|---|
| Kokoro Fast English, the default | ~1 GB | Runs | Runs | Runs |
| Qwen — Fast (0.6B) Quick designed voices | ~3.7 GB | Runs | Runs | Runs |
| Qwen — Higher quality (1.7B) Best expressive designed voices | ~4.7–5 GB | Runs, tight tight | Runs | Runs |
| Qwen — Voice Design Crafting a new character voice | ~3.7–4 GB (base auto-unloaded) | Runs | Runs | Runs |
| Coqui XTTS v2 Pre-designed voices (cloning to come) | ~4 GB | Runs | Runs | Runs |
| IndexTTS-2 · planned Per-line emotion from one designed voice | 12 GB card (likely) | Not on this card | Not on this card | Planned |
| Fish Audio S2-Pro · planned Finest fidelity | 16 GB card | Not on this card | Not on this card | Planned |
| Option | VRAM while analysing | 6 GB | 8 GB | 12–16 GB |
|---|---|---|---|---|
| Cloud (Gemini free tier) Runs in the cloud | none on your GPU | Runs | Runs | Runs |
| CPU-only (local) Runs on your CPU + RAM | none on your GPU | Runs | Runs | Runs |
| Small local (e.g. qwen3.5:4b) Runs on your GPU | ~4 GB | Runs | Runs | Runs |
| Mid local (e.g. llama3.1:8b) Runs on your GPU | ~7 GB | Not on this card | Runs | Runs |
| Large local (e.g. qwen3.5:9b) Runs on your GPU | ~8 GB | Not on this card | Runs, tight tight | Runs |
On cards below about 11 GB, the analysis model and the voices take turns — the analyzer steps aside before speech, so their VRAM doesn't add up. Run analysis in the cloud or on your CPU and it uses no GPU memory at all, leaving the whole card for voices. At 12–16 GB they stay loaded together and run at once. That's why a 6 GB card carries the whole pipeline.
No GPU? You need one for real-time generation — CPU-only works, but it's slow enough that we'd call it a fallback, not a plan. That's the honest limitation of running a full cast on your own machine. If your setup is GPU-free, watch the download page — lightweight engine support is on the roadmap.