Data licenses & attribution
Hanzi Garden is built on open data. Each source below requires — and deserves — attribution.
Dictionary
Dictionary entries come from CC-CEDICT, a community-maintained Chinese–English dictionary published by MDBG, licensed under Creative Commons Attribution-ShareAlike 4.0. Referenced work: CEDICT © 1997–1998 Paul Andrew Denisowski.
Cantonese (Jyutping) readings come from CC-Canto © Pleco Software, licensed under Creative Commons Attribution-ShareAlike 3.0. Words without a CC-Canto reading are composed from its per-character readings. Because the base dictionary is built from Mandarin-centric sources, Cantonese-only words and phrases may be missing.
ShareAlike: the dictionary content as presented here — including our merged Cantonese readings and derived romanizations — remains available under CC BY-SA 4.0 (CC-Canto portions under CC BY-SA 3.0); changes were made to the original datasets.
Stroke order data
Character stroke data comes from Make Me a Hanzi, packaged per character by hanzi-writer-data and animated with hanzi-writer (MIT). The stroke data derives from fonts by Arphic Technology and is used under the Arphic Public License.
The character breakdown (radicals, components, character meanings and etymology hints) comes from the dictionary file of Make Me a Hanzi, which is derived from Unihan and CJKlib and used under the GNU Lesser General Public License (version 3 or later).
Vocabulary lists
- HSK 3.0 word lists (levels 1–6, band 7–9) via krmanik/HSK-3.0: word lists by Mani (CC BY-SA 4.0), based on Pleco's HSK 3.0 list (MIT), with a few OCR errors corrected by us; original standard published by the Chinese Ministry of Education (2021).
- TOCFL word lists (7 bands, 2018 revision) published by SC-TOP, cleaned CSV export via tomcumming/tocfl-word-list (public domain / Unlicense).
Example sentences
Mandarin–English sentence pairs come from Tatoeba, licensed under CC BY 2.0 FR.
Fonts
Generated PDF practice sheets embed a subset of DejaVu Sans for pinyin text (license). Hanzi in PDFs are rendered as vector outlines from the stroke data above, not from a font. Zhuyin in PDFs uses glyph outlines extracted from Noto Sans CJK TC (SIL Open Font License 1.1). Display hanzi use the Noto Serif SC/TC webfonts (© 2017–2024 Adobe, for Google's Noto project; SIL Open Font License 1.1), served from this website itself — your browser never contacts Google Fonts — with system fonts as fallback.
Open-source software
Hanzi Garden is built with free, MIT-licensed open-source software — thank you to the maintainers of Astro, React, hanzi-writer, pdf-lib (and @pdf-lib/fontkit), better-sqlite3, nanostores and ts-fsrs. Audio pronunciation uses your browser's built-in speech synthesis — this site sends your text to no external service.
The Mandarin audio download in the text reader is generated in your browser with the Kokoro-82M v1.1-zh voice model (Apache License 2.0; Chinese voices from a dataset provided by LongMaoData), run by ONNX Runtime Web (MIT). Its Mandarin pronunciation rules are adapted from misaki (Apache License 2.0, itself adapted from PaddleSpeech). The model is served from this site; no external service is contacted.
Modification notice (Apache License 2.0, § 4(b)): the speech model served here is a modified version of the ONNX export onnx-community/Kokoro-82M-v1.1-zh-ONNX (8-bit quantized, commit 6cc0f0d). We replaced its two quantized STFT convolution layers with float layers that use the original model's fp32 STFT weights, because the quantized layers produced silent audio; all other weights are unchanged. The change is applied when the site is built, by scripts/data/patch-tts-model.ts.