Castwright

Any book, performed by a full cast — kept true, kept yours, book after book.

Turn a manuscript into a finished, full-cast audiobook on your own machine — every character in their own voice, consistent across a whole series. Castwright runs locally end-to-end; nothing leaves your computer unless you opt into a cloud analyzer.

Upload .md / .txt / .epub / .pdf / .mobi / .azw3 → an analyzer extracts characters, chapters, and per-sentence speaker tags → assign a voice to each character → generate per-chapter audio → listen, revise, and export to M4B or MP3.

What you get

• Full-cast performance — every character speaks in their own voice; one narrator can't be everyone.
• Series memory — a character keeps the same voice across every book in a series, even when an author renames someone mid-series. (No other tool does this for readers.)
• Designed voices — every character gets a unique voice designed from its persona, kept consistent across the series. (Cloning a voice from your own sample is the next major release.)
• Quality gate — every line is acoustically checked, transcript-verified, and drift-checked before a chapter is assembled, and any that fail are re-recorded automatically; the plainly broken lines never reach your ears.
• On-device by default — analysis and speech run on your machine; cloud is opt-in.
• You own the files — export standard M4B / MP3 / AAC / Opus and keep them. No lock-in.

Features

• Ingest — paste or upload .md, .txt, .epub, .pdf, .mobi, or .azw3. Chapters and character names are extracted automatically; low-confidence speaker tags are surfaced for a quick review pass. DRM-protected files are rejected up front. Re-uploading a book shows a sentence-level diff before anything changes.
• Try a sample — a ready-made demo book (an original full-cast story) loads in one click with its whole cast already designed, so you can generate and hear a full performance before importing anything of your own.
• Analyzer choice — run a fully local model via Ollama (no API key, nothing leaves your machine) or use the Gemini free tier. For long books, an optional two-model pipeline runs cast detection and sentence attribution in parallel to roughly double throughput.
• Voice & cast — per-engine voice catalogues with family grouping, drag- or tap-to-assign, per-character overrides, and sample playback. Audition candidate voices in the profile drawer before committing, and compare two characters side by side — even across books.
• Voice engines — Kokoro (fast, English, runs on a modest GPU), Coqui XTTS v2 (zero-shot voice cloning, optional download), Qwen3-TTS (designs a unique voice per character from a persona and reuses it across the series), and Gemini cloud. A character keeps its voice when you switch engines.
• Generation — per-chapter, resumable, and sticky across navigation; chapters synthesise in a bounded parallel pool. Each chapter opens with its title spoken in the narrator's voice. Open the same book in two tabs and progress stays in sync.
• Revisions & drift — pending-revision review, A/B audition with rollback, a per-chapter revision timeline, and automatic detection (with one-click regeneration) when a chapter's voices drift from the current cast.
• Listening — a built-in player with speed control, markers, a sleep timer, true waveform peaks, resume bookmarks, per-book notes, and one-tap sharing of a 30-second clip or the whole book.
• Library — auto-fetched cover art with a manual cover picker (search / upload / frame), title & author search, tag filters, series grouping, and a portable book bundle for backup or transfer.
• Export — M4B (with cover and chapter markers), AAC/M4A, Opus, MP3 (zip or per-chapter folder), plus LAN download with a QR code. EBU R128 loudness normalisation is on by default.
• Mobile, tablet & companion app — every view is responsive across phone / tablet / desktop and reachable over your home network via HTTPS, with a native Android companion that syncs only what changed and plays offline (background, lock-screen, Bluetooth, Android Auto).

Quickstart

Two ways to install:

• One click, no terminal — Pinokio. Paste the repo URL into the Pinokio browser and click Install. Pinokio provisions its own Python, ffmpeg and voice engine, builds the latest release, and launches the app — nothing to set up by hand. See INSTALL.md → Pinokio.
• Manual — the release zip. Download the latest castwright-vX.Y.Z.zip from Releases, extract it, and follow INSTALL.md. You'll end up with a single npm run start:prod that brings up the server, the voice engine, and the web UI at http://localhost:8080.

Prerequisites (manual install only — Pinokio provisions these itself; full detail and per-OS steps in INSTALL.md):

• Node.js 20.19+
• Python 3.12 (exactly — for the voice engine)
• ffmpeg on PATH
• ~6 GB free disk (TTS weights ~1.1 GB + PyTorch ~2.5 GB + deps)
• A GPU is strongly recommended (TTS on CPU is far slower): NVIDIA on Windows/Linux (PyTorch 2.11.0 installs CUDA-bundled by default; for a specific toolkit like CUDA 12.8 see INSTALL.md), or Apple Silicon (M-series) on macOS — used automatically via Metal, no drivers or setup. AMD GPUs are an experimental preview (ROCm for Qwen/Coqui; Kokoro runs on CPU), auto-detected with a safe CPU fallback — see INSTALL.md. No GPU? It still runs on CPU.

Companion app (Android)

A native Flutter companion pairs to your running server over the home network (HTTPS, cert-pinned), delta-syncs only the chapters that changed, and plays them offline with background, lock-screen, Bluetooth, and Android Auto controls. Pairing is cryptographically self-verified, so no certificate install is needed on the phone.

Each GitHub Release attaches a ready-to-sideload castwright-vX.Y.Z.apk. Server-side pairing setup (LAN HTTPS + an access token) is in INSTALL.md; the app's own build and usage notes are in apps/android/README.md.

Releases

Tagged releases are published to GitHub Releases. Each attaches (all with .sha256 checksums):

• castwright-vX.Y.Z.zip — the platform-independent server bundle (Windows / macOS / Linux); install via INSTALL.md.
• castwright-vX.Y.Z.apk — the sideloadable Android companion app.

An iOS companion build isn't published yet — it lands with app-12 (signed .ipa); the Flutter codebase already stays iOS-ready.

After the first public release, upgrading is one click inside the app (Account → Application updates); see INSTALL.md.

How it's built

Castwright is a Vite + React frontend, a Node/Express server, and a Python (FastAPI) voice engine, with a native Flutter companion app. Building from source, the branching model, and the commit convention are documented in CONTRIBUTING.md.

Documentation

• INSTALL.md — install, configure, and update the app.
• apps/android/README.md — the Android companion.
• CONTRIBUTING.md — building from source, branching, and the commit convention.
• SECURITY.md — supported versions and how to report a vulnerability privately.

License

Castwright is source-available — not OSI open source. The code is licensed under the Functional Source License v1.1 with an Apache-2.0 future grant (FSL-1.1-ALv2, a.k.a. FSL-1.1-Apache-2.0) — see LICENSE. In short: use, modify, and share it for any purpose except building a competing product or service; two years after each release, that release's code becomes Apache-2.0. Leading with this plainly is deliberate — it bars a competing fork from day one while keeping the source fully readable.

• Name & brand — the Castwright name and identity are all rights reserved and not part of this repository; see docs/legal/brand-and-trademarks.md.
• Model weights carry their own upstream licences (NOTICE). Coqui XTTS v2 is non-commercial (CPML) and is therefore download-on-demand, never bundled.
• Privacy & terms — the apps collect nothing and send nothing to us; the full policies live at castwright.ai/legal/privacy and castwright.ai/legal/terms. Castwright is a registered business name (ABN 52 641 247 931) in Australia.

Contributing: issues welcome; PRs by invitation for now (a DCO sign-off and a lightweight CLA apply — see CONTRIBUTING.md).