Context-aware Pronunciation 0.8.3

Context-aware Pronunciation 0.8.3

简体中文

这是 main 候选版本，支持 NVDA 2026.2，并测试 2026.3beta2。
推送 main 与生成本地候选包不等于发布；发布遵循发布流程。

• 按身高名词与量词区分“个／個”：“长个、长个了、长个了没、长不了个、
  长一长个”的“个”选择 gè 四声；不添加“儿／子”。“长个东西”中的
  量词“个”记录轻声，保留原字供声库识别。
• “长＋量词＋名词短语”由实际名词中心语与依存关系选择 zhǎng，
  支持新名词、修饰语、数量、处所、否定和共享谓词形态。
  例如“长了两个黑色的东西、树上长个奇怪的东西、长个公司的招牌”。
  处所分支不再丢弃已成立的量词宾语；时长、年龄和未知续接保持独立判断。
• 固定授权词典中读音明确、包含轻声的完整短词保留词形，避免同音字
  改写破坏声库整词识别。例如“数落、亲家、朋友、明白、麻烦、妈妈、桌子”。
  歧义词和单独后缀不构成轻声证据；实际轻声输出仍依赖声库。
• 保留用户规则、停用规则、逐字模式、原始位置与语音／盲文决策一致性。
  所有处理仍在本机进行，运行时不开网络、不加载大模型、不存储朗读文本。
• 新增 71,280 个构造的量词宾语组合及反例、共享边界和快速生产式检查，
  保持原有短句／8K 性能门槛。Ting-Ting 离线探针使用独立 SDK 拼音四声／
  轻声对照检查实际纯文本输出的完整 PCM，不以音素编号相等推断声调。
  SDK 命令仅用于离线验证，不进入插件输出。

构造语料通过证明的是所测结构覆盖，不是自然语料准确率或穷尽汉语歧义。
声学自动检查不等于人工试听或其他声库认证。
候选安装包为 contextualPronunciation-0.8.3.nvda-addon。

English

• Select fourth-tone gè for clipped stature 个/個, and neutral ge for classifiers
  with overt objects; no missing 子/儿 is inserted.
• Generalize zhǎng through attached classifier noun phrases, preserving actual
  heads and offsets across quantities, modifiers, locations and shared morphology.
  Retain separate duration/age analysis and unknown continuations.
• Preserve complete short lexical neutral words from pinned licensed sources,
  without generalizing from suffixes or choosing unresolved homographs.
• Keep personal controls, spelling mode, item boundaries and shared braille
  annotations. Add constructed variations, counterexamples and unchanged latency
  gates; offline Ting-Ting PCM controls distinguish ge4 from ge5 independently.

Finite constructed coverage and machine acoustic comparisons do not establish
universal ambiguity resolution, human listening or support in every voice.

Context-aware Pronunciation 0.7.9

Context-aware Pronunciation 0.7.9

简体中文

本版包含 PR #4 的口语“懵”纠正和“腾”轻声词处理。需要 NVDA 2026.2 或更高版本；CI 另覆盖 2026.3beta2。

读音调整

• “一脸懵逼、懵了、懵圈、看懵了、懵得说不出话”等口语表达默认选择一声 mēng，也覆盖未逐项列出的新语境。这是口语兼容策略，不表示“懵”的规范字音改为一声。
• “懵懂、懵懵懂懂、懵里懵懂／懵裡懵懂、懵然”等审定书面词保留原字，避免覆盖其三声 měng。保护词表和资料来源见参考资料。需要关闭口语默认时，在“禁用规则 ID”中填写 colloquialMeng；个人 keep 和读音规则仍可覆盖特定语境。
• 为“折腾、倒腾、捣腾、闹腾、掀腾”及相应繁体词形记录轻声 teng5，同时保留完整词形交给语音合成器处理，避免先改成“遮腾、导腾”而破坏整词识别。
• 增加跨词边界保护，例如“曲折／腾挪”“打倒／腾空”不会因相邻字组成“折腾、倒腾”而误标轻声；“折腾空调、倒腾空间、倒腾飞机”仍按完整轻声词处理。

验证范围与限制

新增回归覆盖简繁体、严格模式、用户规则、禁用规则、语音决策与注音的对应关系，以及原文位置和词边界。它们验证规则输出，不构成新增声音的音频实测。

轻声读音记录可供注音接口使用；朗读时保留整词，由当前语音合成器决定实际发音。本版不向所有合成器强制发送轻声音素，也不保证所有声音都能正确读出轻声。书面词保护需要相关字位于同一段输入文本中，不能跨语音命令或分开的语音文本项恢复上下文。

下载 contextualPronunciation-0.7.9.nvda-addon，安装后按提示重启 NVDA。源码包为 contextualPronunciation-0.7.9-source.zip，可用发布附件中的 SHA256SUMS 核验下载。

English

This release includes the colloquial 懵 and neutral-tone 腾 changes from PR #4. Requires NVDA 2026.2 or later; CI additionally covers 2026.3beta2.

Pronunciation changes

• Default colloquial 懵 to first-tone mēng in 一脸懵逼, 懵了, 懵圈, 看懵了 and other contexts, including expressions outside a finite phrase list. This is a colloquial compatibility policy, not a change to the character's normative dictionary reading.
• Preserve 懵懂, 懵懵懂懂, 懵里懵懂 / 懵裡懵懂, 懵然 and other reviewed literary forms to retain third-tone měng. See the references for the protection inventory and sources. Disable the colloquial default with rule ID colloquialMeng; personal keep and reading rules still take precedence for matching contexts.
• Record neutral-tone teng5 for 折腾, 倒腾, 捣腾, 闹腾, 掀腾 and their traditional forms. Keep each complete word for the synthesizer instead of rewriting its first syllable to a homophone such as 遮 or 导.
• Protect boundaries such as 曲折 / 腾挪 and 打倒 / 腾空 from false neutral-tone annotations, while retaining valid complete words in 折腾空调, 倒腾空间 and 倒腾飞机.

Validation scope and limitations

New regressions cover simplified and traditional text, strict mode, user overrides, disabled rules, speech decisions and annotations, original offsets and word boundaries. They validate rule output and do not constitute new voice audio measurements.

The annotation API exposes neutral-tone readings; speech preserves complete words and leaves their acoustic realization to the selected synthesizer. This release does not force neutral-tone phonemes in every engine or guarantee pronunciation across all voices. Literary-word protection requires the relevant characters within one input text item; context is not reconstructed across speech commands or separate text items.

Install contextualPronunciation-0.7.9.nvda-addon and restart NVDA. The matching source bundle is contextualPronunciation-0.7.9-source.zip; verify downloads using the release's SHA256SUMS attachment.

Context-aware Pronunciation 0.7.8

Context-aware Pronunciation 0.7.8

简体中文

本版升级离线中文多音字纠正，并整理用户与贡献者文档。需要 NVDA 2026.2 或更高版本；稳定 API 声明为 2026.2，原生 CI 另覆盖 2026.3beta2。

读音与句法

• 中文纠正启用时始终加载词典与有界语法，移除“启用扩展词组”选项。
• “重装／重裝”默认按重新安装读 chóng zhuāng，保留明确的重型装备等用法保护。
• 扩展“转”的旋转、绕逛与转交义判断，结合名词中心、主宾语、接收者、受益者、目的从句和补语。
• 覆盖“螺丝刀转呀转、引擎会转、发电机转起来”等结构；重复谓词共享论元。“把螺丝刀转给经理”等转交用法保留三声。
• 共用重复构式也用于“盛、量、系”的现有语义框架。补充出差数量／时长、宿的夜数／星宿及“倔强”两个音节的处理。
• 用户规则、keep、原文位置与解析工作预算继续生效。处理始终离线，文档原文保持不变。
• “装盛／裝盛”默认读 chéng；“盛装动作、盛装到容器里”等容纳义使用 chéng，“盛装赴会、身穿盛装”保留原读音。
• “弹……琴／彈……琴”依据动宾关系选择 tán，支持数量、所属关系和嵌套修饰语，例如“弹了123遍琴、弹小明昨天买的那架钢琴”。
• 修复设置中的三个自定义输入框：回车换行、多行粘贴与保存，保留 Tab 导航和 Ctrl+S 应用设置。

文档与验证

用户设置和排查集中在中英文使用指南，安装帮助由同一份指南生成。开发架构、贡献规则和发布流程合并整理；旧过程报告及本地 AI 指令不进入公开包。

发布工作流要求当前提交的两个原生 NVDA CI 目标通过，并验证可复现包与官方清单格式。此引擎版本的本地 Ting-Ting 核验包含 118 组最终朗读、120 组边界和 130 组句子对照。音素／PCM 一致性不等于所有声音和所有语境均正确；未知或有歧义的结构仍可能保留原字。

下载 contextualPronunciation-0.7.8.nvda-addon，安装后按提示重启 NVDA。源码为 contextualPronunciation-0.7.8-source.zip，使用 SHA256SUMS 核验下载。

本次仅发布 GitHub Release。首次官方商店提交尚在等待批准，当前不提交商店更新。

English

This release expands offline Mandarin polyphone correction and consolidates current documentation.

• Always enable dictionary and bounded grammar with Chinese correction; remove the extended-lexicon checkbox.
• Default 重装 / 重裝 to chóng for reinstall, while preserving explicit heavy-equipment expressions.
• Select rotation, roaming and transfer senses from argument heads, recipients, beneficiaries, purpose clauses and verb complements. Add dictionary-derived mechanical/tool candidates and shared arguments for repeated predicates such as 转呀转 and 转了又转.
• Reuse repetition with the existing 盛, 量 and 系 frames; improve counted work trips, 宿 contexts and both syllables of 倔强.
• Keep user overrides, original offsets, offline processing and work limits. Ambiguous or unsupported contexts can still remain unchanged.
• Default 装盛 to chéng and recognize containment action nominals and container goals; preserve attire contexts. Select tán for 弹…琴 from parsed instrument objects, including quantities, possessors and nested modifiers.
• Fix multiline custom-rule input: Enter inserts a newline; Tab navigation and Ctrl+S apply remain available.
• Generate installed help from current bilingual user guides. Exclude local AI instructions and development-session reports from published source and packages.

Requires NVDA 2026.2; native CI additionally tests 2026.3beta2. Publication requires both native targets on this exact commit, reproducible archives and official manifest/schema checks. Local Ting-Ting homophone-anchor tests cover the listed engine version; they do not certify all voices or open-domain accuracy.

Install the .nvda-addon and restart NVDA. The matching source bundle, checksums and CI verification are attached. This is a GitHub release only; Store updates remain on hold pending approval of the first submission.

Context-aware Pronunciation 0.7.4

Context-aware Pronunciation 0.7.4

First public repository release of the offline NVDA Mandarin pronunciation add-on. This release preserves the 0.7.3 runtime, including structural handling of 下边缘, 唱和说, 多行, quantified objects and coordinated subjects, with a queue-stage check after dictionaries and symbols.

• Requires NVDA 2026.2. The stable manifest declares that tested stable API; 2026.3beta2 remains an additional development CI target.
• Adds reproducible public builds, source archives, SHA-256 checksums, maintainer/repository metadata, bilingual user guides and community contribution documents.
• Keeps the extended lexicon and braille enhancements experimental. Supported examples and 1,154 targeted grammar cases do not imply complete Chinese accuracy or support for every voice.
• Runtime processing is offline. Homophones are used only in temporary speech text; Speech Viewer can display them. Original documents are unchanged.

Download contextualPronunciation-0.7.4.nvda-addon to install. contextualPronunciation-0.7.4-source.zip contains corresponding source and licensed data snapshots. Verify downloads against SHA256SUMS.

Public CI runs without proprietary voices. Its success confirms the listed code/data/integration checks, not listening acceptance on all synthesizers. Add-on Store submission and approval are separate from this GitHub Release.

中文：本版为首次公开仓库发布，保留 0.7.3 发音引擎，补齐自动构建、源码包、校验和及社区文档。安装后按 NVDA 提示重启。最低和稳定通道最高测试版本均为 NVDA 2026.2；插件商店是否可见取决于 NV Access 审核及索引状态。

Track the official Add-on Store submission: registration issue #11762 and metadata review PR #11763. Store availability follows NV Access approval and indexing.