独立安装测试版：secv2 + 本地 HD 语音（x86/x64）

独立安装测试版：secv2 + 普通自然语音 + 本地 HD

下载 NaturalVoiceSAPIAdapter_v0.2.9-secv2-hd.1_x86_x64.zip，解压后运行 Installer.exe，保留旁边的 data 文件夹。无需预装上游适配器、旧 HD 桥接器或 TtsApplication，也不需要另外覆盖修复 DLL。

本版内容

• 基于本 fork 的 v0.2.9 和已验证的 Sec-MS-GEC v2 修复，未合入上游 v0.2.10 修复。
• 同时安装 x86/x64 普通本地、Edge/Azure 联网引擎及独立 HD 引擎。
• 支持本机模型的 Xiaoxiao HD、Yunyi HD、Andrew HD、Ava HD。HD 与普通同名音色分别保留、分别选择。
• 公开 COM 预合成服务提供能力探测、异步准备、缓存、配额和取消代次。普通 SAPI 软件可以直接选声音；显式预合成需要客户端接入，不保证所有软件安装后自动无缝。
• 中文统一安装入口、语音设置、版本化升级、失败恢复和卸载恢复。

环境与模型

HD 路径仅验证 x64 Windows 上的系统扩展 2608.11000.0.0。模型与 Windows 系统运行库未打包，使用本机已经取得的模型；可由 Windows 安装语音包，或放入 C:\TTS 的子目录。旧适配器不是先决条件。安装完成后须完全退出并重新打开朗读软件。

晓晓/云逸 HD 已验证中英混读；此处的 Andrew/Ava 本地模型仅声明英语，不把云端多语言能力等同于本地包能力。

已验证与限制

• x86/x64 的 318 个 secv2 向量、实际中英文联网合成及取消通过。
• 普通晓晓/云希、HD 四音色及预合成事件/缓存/代次测试通过。
• 空隔离注册树、禁止读取旧安装位置的独立安装、维护重装与卸载检查通过；升级回滚测试通过。不是全新 Windows 虚拟机上的完整系统回归。
• 未逐一操作回归所有 PDF/EPUB 界面、OCR 与跨页场景，不保证任意负载零停顿。旧 SDK 普通中文末尾书签存在偶发缺失的既有现象，详见验证记录。
• 成功入队的预合成音频须持有 ticket 和 lease 至 END 或 purge 确认；不能入队后立即 Cancel。

许可证与源码

保留上游及第三方许可。EdgeGecV2 为单独可替换的 LGPLv3 DLL，完整对应源码已包含在安装包内，也单列为可选下载，普通使用者不需要另装源码包。

微软 SDK 依其原条款分发，不按本项目 MIT 授权；Embedded Speech 组件的额外分发批准尚未核验。技术验证不构成第三方授权已完成的声明，核对记录见 docs/许可证与依赖.md。发布为测试版不表示全部第三方分发条件已获确认。

安装目录、模型发现、客户端协议、验证范围及构建方法见包内中文说明和仓库 docs。SHA-256 见 SHA256SUMS.txt。

Standalone test release: secv2 + standard natural voices + local HD

Download NaturalVoiceSAPIAdapter_v0.2.9-secv2-hd.1_x86_x64.zip, extract it, and run Installer.exe. Keep the adjacent data folder in place. No prior installation of the upstream adapter, the old HD bridge, or TtsApplication is required, and no separate DLL replacement is needed.

What's included

• Based on this fork's v0.2.9 with the verified Sec-MS-GEC v2 fix. Upstream v0.2.10 fixes are not included.
• Installs both x86 and x64 versions of the standard local engine, the Edge/Azure online engines, and the separate HD engine.
• Supports locally available Xiaoxiao HD, Yunyi HD, Andrew HD, and Ava HD models. HD and standard voices with the same name remain separate choices.
• A public COM preparation service provides capability detection, asynchronous preparation, caching, quotas, and cancellation generations. Ordinary SAPI applications can select voices directly; explicit preparation requires client integration. Installing this release does not guarantee seamless playback in every application.
• A unified Chinese-language installer, voice settings, versioned upgrades, failure recovery, and restoration during uninstallation.

Requirements and models

The HD path has only been verified with system extension 2608.11000.0.0 on x64 Windows. Models and Windows system runtimes are not bundled. Use models already obtained locally: install voice packages through Windows, or place them in subdirectories of C:\TTS. The old adapter is not a prerequisite. After installation, fully exit and reopen your reading application.

Xiaoxiao HD and Yunyi HD have been verified with mixed Chinese/English text. The local Andrew/Ava models used here are only claimed to support English; cloud multilingual capabilities do not establish support in local packages.

Verification and limitations

• Passed 318 secv2 test vectors on x86/x64, actual online synthesis in Chinese and English, and cancellation tests.
• Passed tests for standard Xiaoxiao/Yunxi, all four HD voices, and preparation events, caching, and generations.
• Passed checks using an empty isolated registry tree, standalone installation with access to old installation locations blocked, maintenance reinstallation, and uninstallation. Upgrade rollback tests also passed. This is not a complete system regression test on a fresh Windows virtual machine.
• Not every PDF/EPUB interface, OCR workflow, or cross-page scenario has been manually regression-tested. Zero pauses under arbitrary workloads are not guaranteed. The old SDK has a pre-existing intermittent issue where the final bookmark is missing with standard Chinese voices; see the verification records.
• For prepared audio successfully queued for playback, retain its ticket and lease until END or purge confirmation. Do not call Cancel immediately after queuing it.

Licenses and source code

Upstream and third-party licenses are retained. EdgeGecV2 is a separately replaceable LGPLv3 DLL. Its complete corresponding source is included in the installation archive and is also offered as an optional separate download. Ordinary users do not need to install a separate source package.

Microsoft SDKs are distributed under their original terms, not this project's MIT license. Additional redistribution approval for Embedded Speech components has not yet been verified. Technical validation does not establish that third-party authorization is complete; see docs/许可证与依赖.md for the review record. Publishing a test release does not mean that all third-party redistribution conditions have been confirmed.

See the Chinese-language instructions in the package and the repository's docs for installation directories, model discovery, the client protocol, verification scope, and build instructions. SHA-256 checksums are listed in SHA256SUMS.txt.

Unofficial v0.2.9 Edge voice fix (PR #121)

Unofficial replacement DLLs for the Edge online voice failure discussed in gexgd0419#120.

Fix submitted upstream: gexgd0419#121

Based on v0.2.9; source commit: 37a62a3. Implements the Sec-MS-GEC v2 handshake using edge-tts 7.3.0 as a reference. This is a prerelease test build from this fork, not an official upstream release.

Download the ZIP for both architectures. It contains x64/NaturalVoiceSAPIAdapter.dll and x86/NaturalVoiceSAPIAdapter.dll, replacement instructions, SHA-256 checksums and licenses. Close speech applications, back up the original DLLs, and replace the files in the matching folders of your existing installation. Keep all other dependencies unchanged. If already registered from the same folders, reinstallation is normally unnecessary. 32-bit applications require x86 even on 64-bit Windows.

Validated: Release x64/x86 builds; 318 token comparisons against edge-tts 7.3.0; Edge English on x64/x86; Edge Chinese on x64; local Aria; word/sentence boundaries as supported, bookmarks, visemes and cancellation. ARM64 and Windows XP runtime behavior were not tested.

The modified digest port retains edge-tts's LGPLv3 license. Its license and the original project's MIT license are included. Full corresponding source is available in the source archives for this tag.

ZIP SHA-256: 08da963f056374e3c87c0f4360f695225fba0c52f1592872d2f988f844d286b1