Licenses and third-party components

Running on your device

• 0c selection model: Funmula/0c-sequence-d512 (Hugging Face revision d512-pf512-20260926), released by BigGo’s 0c project, Copyright (c) 2026 Funmula Corp Limited, MIT License. Modified: quantized to INT4 (block 32), not retrained; it runs only on your device, and the original MIT license text ships with the app. huggingface.co/Funmula/0c-sequence-d512 0c.biggo.com
• ONNX Runtime 1.23.2: Copyright (c) Microsoft Corporation, MIT License, with its ThirdPartyNotices, unmodified. It runs the 0c model and the DistilGPT-2 English model on the device. github.com/microsoft/onnxruntime
• libchewing 0.8.5: the Zhuyin engine C library plus its dictionary and word-frequency data, Copyright (c) 1999–2001 Lu-chuan Kung and Kang-pen Chen, 2004–2024 libchewing Core Team and contributors, GNU LGPL 2.1. The code is unmodified; the dictionary is generated from the official phone.cin and tsi.src. Android links it dynamically as a separate libchewing.so; the Windows program contains no libchewing code and only reads, at run time, the dictionary data generated from it in the install folder’s data\chewing, which you can replace. Under the LGPL you may modify this library and data; the corresponding source you need is under “Get the corresponding source” below. codeberg.org/chewing/libchewing
• McBopomofo 3.1 Traditional Chinese lexicon: Copyright (c) 2011–2026 Mengjuei Hsieh et al., MIT License, including the historical libtabe BSD notice. Modified: merged with libchewing word frequencies, variant characters repaired, and a within-word continuation table derived; the libchewing words and readings merged in remain under GNU LGPL 2.1. github.com/openvanilla/McBopomofo
• libopus 1.6.1: Copyright 2001–2023 Xiph.Org, Skype Limited, Octasic, Jean-Marc Valin, Timothy B. Terriberry, CSIRO, Gregory Maxwell, Mark Borgerding, Erik de Castro Lopo, Mozilla, Amazon, BSD 3-Clause License, unmodified, used to compress recordings. opus-codec.org
• Context bigram model: the first source is sample statistics from the 0c selection model Funmula/0c-sequence-d512, revision d512-pf512-20260926, Copyright (c) 2026 Funmula Corp Limited, MIT License. About 1.5 billion characters were generated with the model, segmented with TingKey’s lexicon and counted as word pairs; the generated text itself is not distributed. The second source is libchewing-data chewing_v4 bigram_p50.arpa, Copyright libchewing Core Team (trained by Kan-Ru Chen on the CC-100 corpus), CC BY 4.0. Modified: unigram probabilities interpolated with λ=0.6, mapped to TingKey’s lexicon, mixed at 0.3 for 0c samples and 0.7 for libchewing, pruned to 2,500,000 word pairs and converted to TKBG001 binary format. Under Section 5 of CC BY 4.0, the licensor provides no warranties. huggingface.co/Funmula/0c-sequence-d512 codeberg.org/chewing/libchewing-data/src/commit/e5d726d9c09afd562309b05d9adad29940de5ebb/dict/chewing_v4 creativecommons.org/licenses/by/4.0/
• SymSpell English frequency dictionary: Copyright (c) 2018 Wolf Garbe, MIT License, unmodified. SymSpell built it as the intersection of Google Books Ngram data (Google, CC BY 3.0) and the SCOWL word lists (Copyright 2000–2018 Kevin Atkinson et al., a permissive license that allows commercial use). github.com/wolfgarbe/SymSpell creativecommons.org/licenses/by/3.0/ wordlist.aspell.net

Also in the Android app

From Android 0.9.107, the in-app “Open-source components and licenses” screen matches this page, listing every component with its notice and full license text. Lists in 0.9.106 and earlier are incomplete; this page is authoritative for them.

• OpenCC dictionaries: Copyright BYVoid, Apache License 2.0. Three original dictionary files are unmodified, and a conservative single-character mapping is derived from them for Traditional forms. github.com/BYVoid/OpenCC
• opencc-python-reimplemented 0.1.7 (Yichen Huang), Apache License 2.0: its seven bundled OpenCC dictionary files are used unmodified for script conversion of speech results. Per its NOTICE: “OpenCC-Python, Copyright 2016. This product includes dictionary data and configurations developed by OpenCC.” github.com/yichen0831/opencc-python
• A subset of Noto Sans TC (Sans2.004), renamed TingKey Phonetic, used to draw the Zhuyin symbols on the keys: Copyright 2014–2021 Adobe, with Reserved Font Name “Source”, SIL Open Font License 1.1. Modified: reduced to the Zhuyin symbols and renamed. github.com/notofonts/noto-cjk
• OkHttp 4.12.0 and Okio 3.6.0 (Square, Inc.), Apache License 2.0, unmodified, used for network connections. github.com/square/okhttp www.apache.org/licenses/LICENSE-2.0
• Kotlin standard library 1.9.10 and JetBrains annotations 13.0 (JetBrains s.r.o.), Apache License 2.0, unmodified, bundled with OkHttp. Per Kotlin’s NOTICE: “Kotlin Compiler, Copyright 2010–2023 JetBrains s.r.o and respective authors and developers”. The standard library contains code derived from GWT and Guava (Apache License 2.0) and Boost math functions (Boost Software License 1.0). github.com/JetBrains/kotlin www.boost.org/LICENSE_1_0.txt
• The Public Suffix List bundled in OkHttp (Mozilla Foundation), Mozilla Public License 2.0, unmodified. publicsuffix.org www.mozilla.org/MPL/2.0/
• LLVM libc++ (Android NDK 29): statically embedded in the app’s native libraries, Apache License 2.0 WITH LLVM Exception, unmodified. llvm.org/LICENSE.txt

Also in the Windows app

• DistilGPT-2 (Hugging Face), Apache License 2.0, provides English completion and spelling-suggestion scoring locally on your computer. Modified: exported by TingKey to an ONNX scoring graph and quantized to INT8, not retrained, and not an official export. It is distilled from OpenAI’s GPT-2, which is under the Modified MIT License (Software Copyright (c) 2019 OpenAI). huggingface.co/distilbert/distilgpt2 github.com/openai/gpt-2
• nlohmann/json 3.11.3: Copyright (c) 2013–2022 Niels Lohmann, MIT License. github.com/nlohmann/json
• Microsoft Visual C++ 2015–2022 runtime: installed by Microsoft’s official vc_redist.x64.exe bundled with the installer, opening its installer when needed so the user can accept Microsoft’s license terms; the ONNX Runtime that runs 0c and DistilGPT-2 needs it. learn.microsoft.com/cpp/windows/redistributing-visual-cpp-files
• Inno Setup 6.7.3 (the installer): Copyright (C) 1997–2026 Jordan Russell, portions Copyright (C) 2000–2026 Martijn Laan, Inno Setup License. jrsoftware.org

Models and software on our server

These models run on TingKey’s server and are not included in the app installers.

• Speech recognition: the primary server uses Qwen3-ASR-1.7B and the backup server uses Qwen3-ASR-0.6B (both from the Qwen team), Apache License 2.0. huggingface.co/Qwen/Qwen3-ASR-1.7B-hf huggingface.co/Qwen/Qwen3-ASR-0.6B-hf
• Editing, rewriting, translation and lexicon review: the primary server uses GLM-5.3-Flash (Copyright (c) 2026 Z.AI Co., Ltd, MIT License, run with MLX 4-bit quantization); the backup server uses Qwen3-4B-Instruct-2507 (Qwen team, Apache License 2.0, run with Q4_K_M quantization). huggingface.co/zai-org/GLM-5.3-Flash huggingface.co/Qwen/Qwen3-4B-Instruct-2507
• Script conversion of the raw transcript uses OpenCC (Apache License 2.0). github.com/BYVoid/OpenCC

OpenRouter

• If you choose “OpenRouter: speech + editing” on Android, recordings are processed by Qwen3.8 Omni Flash through your own OpenRouter account, under the terms and pricing of OpenRouter and the model provider.

The website itself

• Text uses your device’s system fonts; the site embeds no web fonts.
• Icons and the logo are drawn by TingKey; screenshots come from TingKey app test screens.
• The site is hosted on Cloudflare Pages; the web speech trial uses Cloudflare Turnstile to prevent abuse, under Cloudflare’s terms of service. www.cloudflare.com/website-terms/

Get the corresponding source

TingKey’s own code for Android and Windows is not published; under the LGPL, we provide the corresponding source of the LGPL components below:

• Corresponding source of the LGPL components in Android 0.9.117: libchewing source code, plus dictionary data derived from libchewing and the scripts that generate it (including instructions for replacing libchewing.so)
• Corresponding source of the LGPL components in Windows 0.1.34: libchewing source code, plus dictionary data derived from libchewing, installed in data\chewing, and the scripts that generate it

Product and project names mentioned here (such as 0c, BigGo, Qwen, ONNX Runtime, OpenRouter and Cloudflare) belong to their owners and are used only to identify sources; mentioning them does not imply partnership or endorsement.