A local, multi-voice audiobook generator: an LLM casts every character in its own voice, kept consistent across a whole series — with voice design, per-line emotion, seven languages, and M4B / Audiobookshelf export. Runs on your own machine; you own the files. Kokoro · Qwen3-TTS.
15
stars
9,541
commits
TypeScript
primary language
Sep 10, 2026
updated
README
Castwright
Any book, performed by a full cast — kept true, kept yours, book after book.
▶ Hear a 90-second sample — the same passage as one narrator, then the full cast: castwright.ai/#demo.
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; with no cloud analyzer key set, nothing leaves your computer.
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 (Qwen3-TTS), kept consistent across the series; Kokoro ships a ready-made catalogue alongside it. (Cloning a voice from your own sample is the next major release.)
- Seven languages — performs English, Russian, Spanish, French, German, Chinese, and Japanese end-to-end, with the manuscript's language detected on import. A cast never crosses languages within a book. (CJK attribution leans on the analyzer model — a local Qwen analyzer is recommended for Chinese and Japanese.)
- Quality-checked automatically — every chapter is gated before it's assembled: acoustic checks (dead air, over-long lines, timing drift), optional word-for-word ASR transcript verification, and an opt-in speaker-fingerprint check that catches a voice drifting out of character even when the words are right. Flagged takes are surfaced on the waveform and re-recorded automatically.
- On-device by default — speech is always local. The analyzer is local too; add a cloud key and it falls back to the cloud automatically unless you switch that off.
- Listen on your phone or tablet — an Android companion app plays from your library over your home network, opening into a two-pane layout on tablets and foldables. Chapters you haven't downloaded play instantly on-network, with no OS certificate install.
- You own the files — export standard M4B (cover and chapter markers), MP3, AAC/M4A or Opus, a native Audiobookshelf export (series metadata,
metadata.json, and a folder cover), and SRT/VTT captions (per sentence or per line, whole-book or per-chapter). EBU R128 loudness normalisation is on by default. No lock-in.
Documentation
The full user guide — installing, uploading a book, casting, generating audio, listening, exporting, and every feature area — lives on the wiki, illustrated with real screenshots.
- New here? Start at Getting Started.
- Installing from the release zip? See INSTALL.md.
- Release history: RELEASE_NOTES.md.
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).
Contributors
Similar repos
Languages
TypeScript
76.0%
JavaScript
12.8%
Python
8.2%
Dart
2.2%
A local, multi-voice audiobook generator: an LLM casts every character in its own voice, kept consistent across a whole series — with voice design, per-line emotion, seven languages, and M4B / Audiobookshelf export. Runs on your own machine; you own the files. Kokoro · Qwen3-TTS.
15
stars
9,541
commits
TypeScript
primary language
Sep 10, 2026
updated
README
Castwright
Any book, performed by a full cast — kept true, kept yours, book after book.
▶ Hear a 90-second sample — the same passage as one narrator, then the full cast: castwright.ai/#demo.
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; with no cloud analyzer key set, nothing leaves your computer.
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 (Qwen3-TTS), kept consistent across the series; Kokoro ships a ready-made catalogue alongside it. (Cloning a voice from your own sample is the next major release.)
- Seven languages — performs English, Russian, Spanish, French, German, Chinese, and Japanese end-to-end, with the manuscript's language detected on import. A cast never crosses languages within a book. (CJK attribution leans on the analyzer model — a local Qwen analyzer is recommended for Chinese and Japanese.)
- Quality-checked automatically — every chapter is gated before it's assembled: acoustic checks (dead air, over-long lines, timing drift), optional word-for-word ASR transcript verification, and an opt-in speaker-fingerprint check that catches a voice drifting out of character even when the words are right. Flagged takes are surfaced on the waveform and re-recorded automatically.
- On-device by default — speech is always local. The analyzer is local too; add a cloud key and it falls back to the cloud automatically unless you switch that off.
- Listen on your phone or tablet — an Android companion app plays from your library over your home network, opening into a two-pane layout on tablets and foldables. Chapters you haven't downloaded play instantly on-network, with no OS certificate install.
- You own the files — export standard M4B (cover and chapter markers), MP3, AAC/M4A or Opus, a native Audiobookshelf export (series metadata,
metadata.json, and a folder cover), and SRT/VTT captions (per sentence or per line, whole-book or per-chapter). EBU R128 loudness normalisation is on by default. No lock-in.
Documentation
The full user guide — installing, uploading a book, casting, generating audio, listening, exporting, and every feature area — lives on the wiki, illustrated with real screenshots.
- New here? Start at Getting Started.
- Installing from the release zip? See INSTALL.md.
- Release history: RELEASE_NOTES.md.
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).
Similar repos
Contributors
Languages
TypeScript
76.0%
JavaScript
12.8%
Python
8.2%
Dart
2.2%