v0.1.6 — Japanese kanji G2P (MeCab + UniDic) & one-click dictionary installer

v0.1.6 — Japanese kanji G2P (MeCab + UniDic) & one-click dictionary installer

日语汉字 G2P（MeCab + UniDic）与一键词典安装

Highlights / 重点

• Japanese kanji lyrics now work. The japanese-mecab converter is a real
  MeCab + UniDic pipeline, exactly like upstream: MeCab segments the lyrics,
  UniDic supplies the whole-word kana readings (pron), and every reading
  candidate goes to the aligner. A kanji word without a reading fails per file
  instead of silently producing wrong phonemes.
  日语汉字歌词可用。 japanese-mecab 现在与上游一致：MeCab 分词，UniDic
  给出整词假名读音（pron），全部读音候选交给对齐器；无读音的汉字词按文件报
  错，不会静默给错误音素。

• Kana lyrics keep working out of the box. Without the dictionary the same
  config id falls back to the kana converter (v0.1.5 behaviour, byte-identical
  output).
  假名歌词照旧开箱即用。 未装词典时同一配置 id 自动回退假名转换器（v0.1.5
  行为，输出逐字节一致）。

• One-click dictionary install in the GUI. The environment panel has a new
  "日语词典" row: downloads with progress and installs to models/unidic/
  automatically.
  GUI 一键安装。 顶栏环境区新增"日语词典"行：带进度下载并自动装到
  models/unidic/。

• unidic-lite-dicdir.zip ships as its own release asset (~49 MB packed,
  ~260 MB unpacked) — deliberately not inside the platform bundles. It is
  probed at models/unidic/ (the converter's unidic_dir kwarg overrides);
  bundle README.txt and docs/usage.md explain it. MeCab is BSD-3 and UniDic
  BSD/GPL/LGPL — see THIRD_PARTY_NOTICES.md.
  词典为独立 release 资产（压缩约 49 MB，解压约 260 MB），刻意不进平台包。
  CLI/GUI 在 models/unidic/ 自动探测（转换器 unidic_dir kwarg 可覆盖）；
  详见 THIRD_PARTY_NOTICES.md 与包内 README.txt。

For OpenUtau users / OpenUtau 用户

The .oudep packages are unchanged: OpenUtau passes phones directly and never
runs G2P, so this release does not rebuild them.
.oudep 无需更新：OpenUtau 直接传入音素、不走 G2P，因此本次不重建 oudep。

Verified / 已验证

可愛いから好きになったなんて with -l ja on the stock model + dictionary:
1 ok, 25 phones, the texts tier keeps the kanji surfaces
(可愛い/から/好き/に/なっ/た/なんて), phones
k a w a i | k a r a | s u k i | n i | n a cl | t a | n a N t e
(促音→cl，ん→N，长音按上游约定丢弃). Stock model + dictionary, -l ja:
1 ok，25 phones，texts 层保留汉字词面；假名回退与 v0.1.5 完全一致。
The GUI one-click install was exercised against this very asset, and the
dictionary archive was downloaded and checked from this release.
GUI 一键安装即从本 release 下载该资产；词典压缩包亦已下载复核。

Thanks / 致谢

MeCab by Taku Kudo / NTT (BSD), UniDic by the UniDic Consortium (BSD),
unidic-lite packaging.

v0.1.5

What's Changed

• fix(ui): resolve file and folder picker buttons not adding inputs by @Xiantaidu in #6
• 2pass: keep the phrase line, and label the gaps by @KakaruHayate in #7
• breath: port FoxBreatheLabeler as a second AP detector by @KakaruHayate in #8
• breath: FoxBreatheLabeler as the default AP detector (re-opens #8) by @KakaruHayate in #9

New Contributors

• @Xiantaidu made their first contribution in #6

Full Changelog: v0.1.4...v0.1.5

v0.1.4

Full Changelog: v0.1.3...v0.1.4

v0.1.3

tifa.cpp v0.1.3

两件事：PFML 1.0 一致性补齐，以及 word 层语言解析（三态）。命令行对齐结果不受影响。

新增

• 语言三态。PFML 1.0 的语言有三种状态，此前被压成了一个字符串：

  PFML 写法	含义	v0.1.2	v0.1.3
  不写 language	继承外层 scope	报 ""	继承最近 scope 的语言
  language=""	无语言（清空）	""	""
  language-kind="any"	任意语言	接受后忽略	独立状态 language_is_any
  language="zh"	该标签	"zh"	"zh"

• 公开 API 相应扩展：G2PWord 与 G2PWordCandidates 新增
  bool language_is_any；G2PWord::language 现在返回解析后的值。

• to_pfml 把 ANY 写回 language-kind="any"（且不写 language 属性），往返保持一致。

修正（PFML 1.0 一致性）

输入	规范	v0.1.2
<!-- 注释 -->	忽略	解析报错
<![CDATA[重]]>	普通字符数据	解析报错
未知属性（如 bogus="1"）	报错	静默忽略

现在注释与 CDATA 可用，未知属性按规范报错（拼错属性名不会再被静默吞掉）。

兼容性

• 命令行对齐结果零变化：8 个覆盖各语言状态的输入（普通、scope 内、异语言
  scope、多语言 -l zh,yue、显式多标签属性、裸音素、清空、ANY）与 v0.1.2
  产出逐字节相同的 TextGrid。
• 显式多语言属性（language="zh,yue"）仍按首个标签解析，与之前一致。
• 其余修正只影响此前报错或标错的输入。

给 MaxLabel（PFML 编辑器）

• G2PWord::language 的语义变了：对于省略 language 的 word，现在返回外层
  scope 的语言，以前是空串。如果编辑器把空串当作"未确定"来显示，需要改用
  新规则：空串 = 无语言（或已清空），language_is_any = ANY。
• G2PWord::language_is_any 是新字段，G2PWordCandidates 上也有。
• 依赖 pin 可以从 v0.1.2 移到 v0.1.3。
• 验收清单：KakaruHayate/MaxLabel#1 正文里的表格；编辑器侧的
  tests/test_pfml_conformance.cpp 是对应落点。

下载

与 v0.1.2 相同：三平台 GUI 自包含包（tifa-label-<平台>）+ 三平台 CLI 包
（tifa-cli-<平台>-{full,q4}）+ BreathLab 模型源文件。用法见包内 USAGE.md。
命令行务必带 -l（如 -l zh）。

CUDA 包由独立的 cuda workflow 编译，完成后自动补挂到本 release
（tifa-cuda-<平台>-{full,q4}.tar.gz）。

---

English

Two things: PFML 1.0 conformance, and word-level language resolution
(three states). Alignment output is unaffected.

New — a word's language now resolves against its enclosing scope: omitted
inherits, language="" clears, language-kind="any" is a distinct state.
G2PWord and G2PWordCandidates gain bool language_is_any, and
G2PWord::language now reports the resolved value. to_pfml writes ANY
back as language-kind="any".

Fixed — comments (<!-- -->) and CDATA are accepted instead of failing to
parse, and unknown attributes are errors rather than being silently ignored.

Compatible — 8 inputs covering every language state produce byte-identical
TextGrids against v0.1.2, and an explicit multi-language attribute still
resolves through its first tag.

For MaxLabel — G2PWord::language now reports the enclosing scope's
language for a word that omits it (it used to be empty), and
language_is_any is new. The pin can move to v0.1.3; the acceptance table is
in KakaruHayate/MaxLabel#1.

Same packaging as v0.1.2. Pass -l (e.g. -l zh) on the command line. CUDA
packages are appended to this release when the separate cuda build finishes.

v0.1.2

tifa.cpp v0.1.2

这一版把 G2P / PFML 抽成可以脱离 ggml 单独链接的库，并补全了 PFML 1.0 的完整发音树。

新增

• 公开 G2P API（include/tifa_ggml/g2p.h）：文本/PFML 转词（convert /
  convert_pfml）、写回 PFML（to_pfml，带 round-trip 保证）、纯语法校验
  （validate_pfml，不需要模型）、候选读音（candidates）、音素校验
  （resolve_phoneme）。
• ggml-free 构建：-DTIFA_GGML_BUILD_MODEL=OFF 只构建 tifa_ggml_g2p，
  完全不拉取 ggml，适合"对齐器之前"的标注编辑器（如
  MaxLabel）：

  set(TIFA_GGML_BUILD_MODEL OFF)
  add_subdirectory(tifa.cpp)
  target_link_libraries(your_editor PRIVATE tifa_ggml::g2p)

• 完整 PFML 1.0 发音树：<reading> / <path> / <group> 嵌套，<phoneme>
  的 language / symbol 属性，以及省略容器只写音素的片段。

修正

• 标签填充只在属性缺失时生效：text="" / script="" 是作者显式给的空值，
  往返后保持不变（此前会被音素"填"上，破坏 round-trip 保证）。
• <word><phoneme>a</phoneme><phoneme>b</phoneme></word> 此前解析成空序列直接
  报错，现在正常。
• <phoneme language="zh">ong</phoneme> 现在解析为 zh/ong（此前忽略 language）。
• <word text="重" ...> 现在认 text 属性（此前忽略，用音素拼接当词形）。

兼容性

<word phonemes=...>、裸 <phoneme> 串、<scope> 这三种经典写法行为完全不变
（同一批输入对比，输出的 TextGrid 逐字节相同）。上面几条"修正"只影响此前报错或
标错的输入。

下载

与 v0.1.1 相同的布局：三平台 GUI 自包含包（tifa-label-<平台>）+ 三平台
CLI 包（tifa-cli-<平台>-{full,q4}）+ BreathLab 模型源文件。用法见包内
USAGE.md（中英双语）。命令行务必带 -l（如 -l zh）。

CUDA 包由独立的 cuda workflow 编译，完成后自动补挂到本 release
（tifa-cuda-<平台>-{full,q4}.tar.gz）。

---

English

G2P and PFML are now a library you can link without ggml, and the PFML 1.0
pronunciation tree is complete.

New — a public API in include/tifa_ggml/g2p.h (convert, convert_pfml,
to_pfml with a round-trip guarantee, model-free validate_pfml, candidates,
resolve_phoneme); a ggml-free build (-DTIFA_GGML_BUILD_MODEL=OFF) that
fetches no ggml at all, for editors that sit before the aligner; and full
<reading>/<path>/<group> trees with language/symbol phoneme
attributes.

Fixed — label filling now applies only when an attribute is absent, so an
explicit text=""/script="" survives a round trip; <word> with <phoneme>
children no longer parses to an empty sequence; phoneme-level language and the
word-level text attribute are honoured.

Compatible — the three classic forms (<word phonemes=...>, bare
<phoneme> runs, <scope>) are byte-for-byte unchanged; the fixes only affect
inputs that previously errored or were mislabelled.

Same packaging as v0.1.1. Pass -l (e.g. -l zh) on the command line. CUDA
packages are appended to this release when the separate cuda build finishes.

tifa-ggml OpenUtau dependency packages (q4)

OpenUtau dependency packages (.oudep) for the tifa.cpp CLI, used to extract phoneme timing from a recording.

Three q4 packages, one per platform:

• tifa-ggml-windows-x64-q4.oudep
• tifa-ggml-linux-x64-q4.oudep
• tifa-ggml-macos-arm64-q4.oudep

Built from the v0.1.5 CLI bundles with Misc/tifa-oudep/package.py (OpenUtau repo). Install by opening the .oudep file in OpenUtau (it extracts to Dependencies/tifa-ggml).

Each package carries the Q4_0 aligner, the syllable dictionaries and both AP detectors: models/breath-fbl-q4_0.gguf (FoxBreatheLabeler, the default) and models/breath-v5-24k-f16.gguf (BreathLab, the alternative).

v0.1.1

tifa.cpp v0.1.1

相对 v0.1.0 的修正与新增。

这一版改了什么

• 呼吸模型的归属修正：这个阶段之前被标成 FBL，实际上用的是 BreathLab
  模型（作者 Xiantaidu）。全仓的文档、界面、
  命令行提示都已更正并注明出处。
• 诊断 JSON 默认导出：<名称>.diagnosis.json 里是 agreement /
  confidence / determinacy / monotonicity。按它排序，只需要人工校对最差
  的 10%。不想要就传 --output-formats textgrid（界面里取消勾选）。
• 图形工具支持三平台：Linux、macOS 也出 GUI 包了，不再只有 Windows。
• 每个平台一个自包含压缩包：GUI + 引擎 + 运行库 + 模型 + 词典 + 说明全在
  一个包里，下载一个就能用，不需要自己拼装。

下载哪个

包	说明
tifa-label-windows-x64.zip	Windows 图形标注工具，解压即用
tifa-label-linux-x64.tar.gz	Linux 图形标注工具（需 libnss3 libatk-bridge2.0-0 libgtk-3-0 libgbm1 libasound2）
tifa-label-macos-arm64.tar.gz	macOS 图形标注工具（未签名，首次右键 →「打开」）
tifa-cli-<平台>-full.tar.gz	命令行版，F16 全精度模型
tifa-cli-<平台>-q4.tar.gz	命令行版，Q4_0 最小模型

<平台> = windows-x64 / linux-x64 / macos-arm64。包内已含模型、词典与
运行库。命令行务必带 -l（如 -l zh）——模型是多语言的、音素表按语言限定。

breathLab_models_dml.zip 是呼吸模型的上游源文件，供自行转换
（scripts/convert_breath_to_gguf.py）。

CUDA 包由独立的 cuda workflow 编译，完成后会自动补挂到本 release 上
（tifa-cuda-<平台>-full/-q4.tar.gz）。

---

English

Fixes and additions since v0.1.0.

• Breath model attribution corrected: the stage was labelled FBL but the
  model is BreathLab, by Xiantaidu. Docs,
  UI and CLI text all say so now.
• The diagnosis JSON is exported by default (<name>.diagnosis.json with
  agreement / confidence / determinacy / monotonicity) — sort by it and
  proof-read only the worst ~10%. Opt out with --output-formats textgrid.
• GUI builds for Linux and macOS, not just Windows.
• One self-contained archive per platform: app + engine + runtime + models
  • dictionaries + guide. Nothing to assemble by hand.

tifa-label-<platform>.{zip,tar.gz} is the GUI; tifa-cli-<platform>-{full,q4}.tar.gz
is the CLI (F16 and Q4_0 weights). <platform> is windows-x64, linux-x64 or
macos-arm64. Pass -l (e.g. -l zh) on the command line — the model is
multilingual and its symbols are language-qualified. Linux GUI needs
libnss3 libatk-bridge2.0-0 libgtk-3-0 libgbm1 libasound2; the macOS build is
unsigned (right-click > Open).

CUDA packages are built by the separate cuda workflow and are attached to this
release once that build finishes.

v0.1.0

tifa.cpp v0.1.0

openvpi/TIFA 的 C++ 原生（ggml）推理实现——用于歌声数据集制作的多语言强制对齐器。运行时不需要 Python。

这个版本有什么

• 原生推理：CPU / Vulkan / Metal / CUDA；边界与参考实现一致（浮点差 ≤ 1e-3）
• TIFA Label：Windows 图形标注工具；模型与引擎按默认相对路径自动导入，无需配置
• 完整数据集流程：TIFA 对齐 → FBL 呼吸检测（AP/SP）合并 → 2PASS 重对齐
• 多语言 G2P：中文 / 粤语 / 日语 / 英文（含 LSTM OOV 推理），支持 PFML 输入
• 两种精度：F16 全量与 Q4_0 最小，每个平台各一份
• 对齐速度（RTX 2070 / Vulkan / F16）：约 0.15 s 每文件，≈ 42× 实时

下载哪个

包	说明
tifa-label-windows-x64.zip	Windows 图形标注工具，解压即用
tifa-cli-<平台>-full.tar.gz	命令行版，F16 全精度模型
tifa-cli-<平台>-q4.tar.gz	命令行版，Q4_0 最小模型

<平台> = windows-x64 / linux-x64 / macos-arm64。包内已含模型、词典与 MSVC 运行时，开箱即用；用法见包内 USAGE.md（中英双语）。注意命令行必须带 -l（如 -l zh），模型是多语言的、音素表按语言限定。

breathLab_models_dml.zip 是呼吸（FBL）模型的上游源文件，供自行转换用（scripts/convert_breath_to_gguf.py）。

CUDA 包由独立的 cuda workflow 构建，编译完成后会补挂到本 release 上。

---

English

Native C++ (ggml) inference for openvpi/TIFA, the multilingual forced aligner used to build singing-voice datasets — no Python at runtime. Runs on CPU / Vulkan / Metal / CUDA, with boundaries matching the reference implementation (≤ 1e-3 float drift).

Pipeline: TIFA align → FBL breath AP/SP detection merged into the phones tier → 2PASS re-align.

Packages: tifa-label-windows-x64.zip is the Windows GUI annotator (unpack and run; it picks up the bundled models automatically). tifa-cli-<platform>-{full,q4}.tar.gz for windows-x64 / linux-x64 / macos-arm64 — models, G2P dictionaries and the MSVC runtime are included, so they work out of the box; see USAGE.md inside. Pass -l (e.g. -l zh) — the model is multilingual and its symbols are language-qualified.

breathLab_models_dml.zip is the upstream source for the breath (FBL) model, for converting yourself with scripts/convert_breath_to_gguf.py.

CUDA packages are built by the separate cuda workflow and are attached to this release once that build finishes.

edge build

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge

Full Changelog: https://github.com/KakaruHayate/tifa.cpp/commits/edge