Good words have good sources.

Original work

ToneWick (formerly Lantern) is an independent self-hosted reader. Application code, sample texts and the small starter glossary are original and MIT-licensed.

Dictionary

ToneWick uses CC-CEDICT, by the CC-CEDICT contributors and distributed by MDBG. The installer preserves the downloaded snapshot's licence (CC BY-SA 4.0 or 3.0 as actually declared), source header and attribution. ToneWick reformats the source into JSON; that dataset retains its applicable share-alike terms. The exact snapshot metadata and retained notice are in the installation's data/provenance.json and data/CC-CEDICT-NOTICE.txt after preparation. Starter installations contain only the original small glossary.

Language engines

ToneWick uses locally hosted pinyin-pro and OpenCC-JS. Their exact versions/hashes are recorded in data/provenance.json, and upstream notices are retained under vendor/licenses/. Automatic pinyin, script conversion and segmentation are aids, not infallible authorities. Taiwan readings, names and polyphonic words deserve review.

PDF text extraction

ToneWick uses locally hosted Mozilla PDF.js with its character maps. The package version and asset hashes are recorded in data/pdf-provenance.json and upstream notices in vendor/pdfjs/licenses. The PDF component does not include PDF font programs or an OCR engine. PDFs are parsed in your browser; scanned pages need a separate OCR step. Extraction is best-effort and should be reviewed.

Taiwan and HSK learning levels

TBCL is a Taiwan language framework; TOCFL is a proficiency test. They are not alternate labels for HSK. Grades in this reader come only from the active authorized list, identified by system, source and edition in the levels panel. HSK 2.0 and HSK 3.0 remain separate; see the official HSK site. No list is bundled without established redistribution rights. A word's listed grade is not an assessment of its reader's proficiency.

Official NAER reference and vocabulary download page

Colour, without overclaiming

Eight palettes use analogous hues and restrained complementary accents. No palette is guaranteed scientifically pleasing to everybody. Colour preference and harmony are distinct, context-sensitive judgements. See Schloss & Palmer (2011), Aesthetic response to color combinations: preference, harmony, and similarity. Colour coding is optional; text labels remain available.

AI practice stories

AI story generation is disabled on this deployment. Ordinary reading and saved vocabulary remain local.

Your device, your text

ToneWick self-hosts ten open-source font families via Fontsource, including Huninn (jf open 粉圓) by justfont and Cactus Classical Serif (仙人掌明體), both under the SIL Open Font License 1.1; the original source ZIP contains no font binaries. Exact versions and notices are retained in data/fonts-provenance.json and each installed font’s OFL.txt. The live gallery distinguishes loaded fonts from device fallbacks. Local-only OS font choices remain available. Missing characters can use fallback faces; font choice does not convert the text. Text analysis takes place locally. Browser speech may use a network voice, disclosed before playback. Saved books are browser-local and not encrypted or synchronized. Keep backups. Browser speech boundary support varies; see the documented boundary event limitations.

ToneWick voice and optional motion

The optional self-hosted voice uses MeloTTS Chinese (MIT) through sherpa-onnx. Installed model/package licences remain with the server assets. Optional character alignment uses an Apache-2.0 Chinese acoustic model with original bounded alignment code. No Pleco voice or commercial audio is copied.

The optional tone-motion experiment is motivated by pitch-gesture research; that study did not test this application. Motion is not acoustic pitch tracking or proven memory training. It can be disabled and follows reduced-motion accessibility principles.

Natural narration and Chinese wordmark

The original starter stories now include synthetic connected Mandarin narration using sherpa-onnx and MeloTTS Chinese (MIT). Approximate word/character timing uses the Chinese Wav2Vec2 CTC model (Apache 2.0), with a still passage marker where alignment is uncertain and motion only for individually timed characters. These are synthetic recordings, not copied human narration.

聲芯 (shēng xīn) is ToneWick’s coined Chinese name: sound + wick. The Ministry of Education dictionary includes the lamp-wick meaning of 芯. The brush wordmark is original AI-generated lettering, informed by traditional running-script references. No historic artwork or signature is reproduced. Detailed provenance and calligraphy references are retained in the repository’s BRAND document.

Official Taiwan vocabulary

TBCL headwords and original level codes come from the National Academy for Educational Research, April 2025 word list. Reused with attribution under the NAER public reuse declaration. ToneWick expands explicit written variants and removes sense-number suffixes; starred sublevels remain unchanged. Data notice. No agency endorsement is implied.

Audio and visual learning

Liu et al. (2011) studied visual tone contours alongside pinyin. This supports the design principle of coordinating pitch direction, spelling and sound. It does not establish a memory benefit for ToneWick’s animated characters. Alignment also uses a CC-CEDICT-derived syllable map; rare recovered positions are explicitly approximate.

English translation and news reading

Free passage translation uses Helsinki-NLP OPUS-MT Chinese–English, by the Helsinki-NLP/OPUS-MT contributors, licensed CC BY 4.0. ToneWick converts the pinned model to int8 with CTranslate2 and tokenizes with SentencePiece. Model revision, transformations and hashes are recorded in the private runtime’s model-lock.json. There is no endorsement by the authors. Original English companions to ToneWick’s starter stories are MIT-licensed with the application.

Local name hints use jieba-wasm and jieba-rs, under their MIT licences. Bundled licence. Idiom and known-name classifications derive from the attributed CC-CEDICT entries above. Possible name labels are heuristic reading aids, not verified identities or calibrated confidence scores. The article and all such analysis stay in the browser.

Stroke order

Stroke-order diagrams, animation and tracing use Hanzi Writer 3.7.3 by David Chanin (MIT) with character data from hanzi-writer-data 2.0.1, derived from Make Me a Hanzi by Shaunak Kishore and the Arphic PL KaitiM GB and UKai fonts, © 1999 Arphic Technology Co., Ltd. The data files are served verbatim from this site under the Arphic Public License; see also the upstream notice. Where a character is missing from the data, no stroke order is guessed.