mimi v1.5.17

English

• Android now has the same seven interface languages as desktop: English, Simplified Chinese, Traditional Chinese, Japanese, Korean, French and German, plus Follow system. The selection is saved independently of subtitle languages.
• Android subtitle size has a labeled slider. Font size, colors, opacity and history settings update the existing floating subtitles immediately. An empty compact window no longer leaves a background over the playing app.
• Optional Android explanations move into nearby help icons. Hover or keyboard focus shows the explanation; tapping opens readable, scrollable details. Large system fonts wrap controls and provider names. Home and preview copy uses plain functional labels and sample sentences.
• New desktop installations start without a preset speech configuration and guide you to add a service. Existing configurations and saved credentials are preserved.
• Windows application audio selection shows local application icons when available. Subtitle position locking is grouped with the other display controls in desktop Settings → Subtitles.

Validation: 1,257 Rust tests passed (2 ignored), 2,019 frontend tests passed, and both Android variants passed 148 JVM tests each, including actual JNI. Android API 32/35 checks cover language persistence and floating subtitle behavior; API 35 checks also cover all seven languages with 200% system font, native hover, focus, help dialogs and independent translation forms. A short Bilibili playback check delivered English/Chinese captions through Alibaba Cloud and stopped normally. Captions visibly lag the video's embedded subtitles; accuracy and end-to-end latency have not been accepted or measured. Physical-device behavior, sustained sessions, Intel Mac runtime and installed updater transitions remain unverified. Native Before/After and recordings.

Desktop updates are available in Settings → General. Portable ZIP and .deb packages require manual installation; Android users install the new APK over the existing app. Some Android playback apps block audio capture. macOS retains its existing fixed signing identity and is not Apple-notarized. The existing Baidu silence-session limitation remains unresolved.

Full changelog

中文

• Android 补齐桌面端的七种界面语言：英语、简体中文、繁体中文、日语、韩语、法语和德语，并支持跟随系统。界面语言独立保存，不影响字幕语言。
• Android 字号改为带当前数值的滑杆。字号、颜色、透明度和历史设置即时应用到已有浮窗；没有内容的紧凑浮窗不再留下一块背景遮住播放应用。
• Android 的可选说明收进附近的帮助图标。悬停或键盘聚焦显示说明，点按打开可滚动详情；大系统字号下，控件和服务名称按空间换行。首页和预览改用普通功能名称与示例句。
• 桌面端首次安装不再预建语音配置，并引导添加服务。已有配置和保存的凭据不变。
• Windows 应用音频选择在可读取时显示应用本地图标。桌面端设置 → 字幕中的位置锁定与其他显示控件放在一起。

验证：Rust 1,257 项通过（2 项忽略）、前端 2,019 项通过，Android 两种构建各有 148 项 JVM 测试通过，包含真实 JNI。Android API 32／35 验证了语言持久化与浮窗行为；API 35 还验证了七种语言、200% 系统字号、原生悬停、聚焦、帮助弹窗和独立文字翻译表单。短时 B 站实播通过阿里云显示了英中字幕，并正常停止。字幕相对视频内嵌字幕有可见延迟，准确率和端到端延时尚未验收或测量。真机、持续长会话、Intel Mac 运行和已安装版本的更新过程仍未验证。原生 Before/After 与录屏。

桌面端可在设置 → 通用中更新；绿色 ZIP 与 .deb 包需手动安装，Android 用户使用新 APK 覆盖安装。部分 Android 播放应用会禁止声音采集。macOS 保持现有固定签名身份，未经过 Apple 公证。百度静音阶段会话失败的既有限制仍未解决。

完整更新记录

mimi v1.5.16

English

• Choose Qwen-MT Lite, Flash or Plus for built-in Alibaba text translation in Settings → Speech & Translation → your Alibaba configuration → Translation model. Short labels describe each model's focus and streaming behavior in all seven desktop interface languages.
• Save the choice independently for each configuration. New and older configurations without a saved choice default to Lite; configurations with a saved choice keep it. Stop subtitles before switching, then restart to use the selected model.
• Use the selected model for both subtitle translation and the text connection check. Lite and Flash stream text; Plus returns the complete translation. Audio 3.0 recognition, credentials, preview pacing and the current language choices remain unchanged.

This selector is available on desktop. Android retains its existing Lite default and does not add a model selector in this release.

Validation: 1,253 Rust tests passed (2 ignored), 2,005 frontend tests passed, and shared-core/JNI, formatting, strict Clippy, lint, typecheck and production build passed. A signed macOS development app verified model selection, saving and the option labels; seven-language browser layout checks passed 84 cases. A bounded comparison completed 36 real text requests across the three models using six synthetic English/Japanese samples. It does not establish end-to-end subtitle latency or a general quality ranking. Windows/Linux native model switching, Intel Mac runtime, Android hardware and installed updates remain unverified.

Desktop updates are available in Settings → General. Portable ZIP and .deb packages require manual installation. macOS retains its existing fixed signing identity and is not Apple-notarized. The existing Baidu silence-session limitation remains unresolved.

Full changelog

中文

• 在设置 → 语音与翻译 → 阿里云配置 → 翻译模型中，为内置文字翻译选择 Qwen-MT Lite、Flash 或 Plus。七种桌面界面语言均提供简短括号说明，标明各模型的定位和流式返回方式。
• 每个配置独立保存模型。新建配置和没有保存过模型选择的旧配置默认使用 Lite；已有选择的配置保持所选模型。切换前先停止字幕，再重新启动以使用新模型。
• 字幕翻译与文字连接检查均使用所选模型。Lite、Flash 流式出字，Plus 整段返回。Audio 3.0 识别、凭据、预览节奏和当前语言选项保持不变。

模型选择器仅提供于桌面端。Android 本次保持原有 Lite 默认值，不新增模型选择器。

验证：Rust 1,253 项通过（2 项忽略）、前端 2,005 项通过，共享核心／JNI、格式、严格 Clippy、lint、类型检查和生产构建均通过。签名 macOS 开发版验证了模型切换、保存及选项说明；七语言浏览器布局检查共 84 组通过。使用六组英／日合成文字，对三个模型完成了 36 次真实文字请求。这不代表字幕端到端延迟或通用质量排名。Windows／Linux 原生模型切换、Intel Mac 运行、Android 真机与已安装版本更新仍未验证。

桌面端可在设置 → 通用中更新；绿色 ZIP 与 .deb 包需手动安装。macOS 保持现有固定签名身份，未经过 Apple 公证。百度静音阶段会话失败的既有限制仍未解决。

完整更新记录

mimi v1.5.15

English

• Add Apple Translation as an independent on-device text translator for Apple Speech, Alibaba recognition and custom recognition profiles. It requires no translation API key or text proxy; recognition continues to use the selected recognition service.
• Show the selected source/target pair's native availability and readiness. Use Download or enable languages in Settings to prepare the pair through Apple's confirmation interface. Translation resources are separate from speech recognition resources. Connection checks and subtitle sessions never request downloads.
• Simplify Apple Speech language setup to one Recognition Language selector. Choose a language, then use Download and use for missing resources or Set recognition language for downloaded resources. Downloading and applying share one action; a failed save keeps the previous active language. Applying a language in an inactive profile activates that profile and language together. Recognition checks follow the visible selection.
• Distinguish Apple Speech resources that are missing, downloading or temporarily unavailable. Refresh resource status after preparation failures and show recovery guidance for system speech-service errors. Existing system resources can become available immediately; this does not mean a full model was downloaded in that time.
• Preserve saved translator credentials when switching to Apple Translation, keep recognition-key edits independent, and retain original-only subtitles and explicit same-language passthrough.
• Complete model-specific language choices for Gemini, Azure, xAI, Baidu, Tencent, Volcano, DeepL and DeepLX. Preserve documented automatic detection, distinguish Chinese/English mixed modes, and filter source-dependent translation directions. Settings, tray and floating controls share the same choices and saved selection. Target languages are selectable in tray and floating controls, including integrated services such as Gemini; unsupported directions are rejected before saving. Align recognition and target fields in settings.
• Clarify that Mimi's current Google Gemini Live integration detects the input language automatically; all 78 documented translation targets are selectable. Other services retain their own model-specific limits.
• Synchronize Android provider catalogs and protocol mappings, including its separate Alibaba realtime translation and recognition models. Expanded configurable language choices for custom translators do not guarantee support from a user-supplied model.
• Let each desktop configuration explicitly remember its recognition and translation languages. Selecting or reapplying that configuration restores the saved pair; temporary language changes leave it intact. Configurations without a saved pair retain the existing behavior. Android does not yet offer this editor.
• Allow language changes from Settings while listening or paused, preserving the paused state. Restore switching from original-only subtitles back to Apple Translation without a false missing-credentials error. Configuration editing and language preparation still require stopping subtitles.
• Add a close button to the floating subtitle toolbar and collapsed bar. It stops the subtitle session and hides its windows. Show provider logos in tray and floating configuration pickers, and avoid unnecessary Settings redraws on each subtitle update.
• Show saved language combinations below configuration names in tray and floating menus. Clicking the subtitle header’s service information opens the current configuration details directly.
• Add Traditional Chinese, Korean, French and German desktop interfaces and README guides.
• Accept valid empty Volcano subtitle events and assemble bounded incremental previews on desktop and Android. The original Japanese disconnect reported in #184 remains un-reproduced.
• Keep compact subtitle drag hints clear of nearby controls, and localize Volcano language-error recovery actions.
• Add bilingual provider setup and model guides.

Apple Translation is available only on supported Apple silicon Macs running macOS 26 or later and requires an explicit source language for translation. Actual language support and readiness come from the system. Intel Macs, Windows, Linux and Android do not offer Apple Translation. Cloud recognition still sends audio to its selected service.

Desktop updates are available in Settings → General; portable ZIP and .deb packages require manual installation. macOS retains its existing fixed signing identity and is not Apple-notarized. This release does not add Apple Translation to the Android APK.

Validation: release preparation checks passed: 1,248 Rust tests passed (2 ignored), 2,000 frontend tests passed, and shared-core/JNI, formatting, strict Clippy, lint, typecheck and production build passed. Android Debug/Release each passed 145 tests including host JNI, plus both lint checks. Unified integration CI passed for macOS, Windows, Linux and Windows ARM64 compilation. Seven-language browser layout checks and focused automated regressions passed. Signed macOS UI-only checks covered interface switching, current-configuration navigation, saved-pair wrapping, and empty/paused/collapsed subtitle controls; these checks do not establish real capture or provider quality. Earlier signed macOS development runs passed Apple text checks for English → Chinese and English → French, French recognition with original-only subtitles, and a short synthetic Japanese → Chinese Apple Speech + Apple Translation subtitle session through system audio. Pause/stop and language switching were also checked within the recorded scope. Resource preparation can reuse system assets; a cold first-time download and same-process recovery from a native resource failure remain unverified. Language catalogs have shared desktop/Android request coverage; expanded cloud language pairs were not individually tested with live audio. Windows/Linux native behavior, Intel Mac runtime, Android hardware, installed updates and sustained provider stability remain unverified. See the integration record for tested revisions and limits.

Known limitation: Baidu sessions may still fail during silence while audio continues to be sent. The cause remains unresolved, and retry does not guarantee sustained recovery.

Full changelog

中文

• 新增 Apple 翻译，可作为 Apple Speech、阿里云识别和自定义识别配置的独立本地文字翻译服务。不需要文字翻译 API Key 或文字代理；语音识别仍使用所选的识别服务。
• 显示当前识别语言与翻译语言是否受系统支持、是否已就绪。在设置中点击「下载或启用语言包」，通过 Apple 的确认界面准备这组语言。文字翻译资源与语音识别资源分开；检查连接和启动字幕不会触发下载。
• Apple Speech 只保留一个「识别语言」选择器。选好语言后，缺少资源时点击「下载并使用」，已有资源时点击「设为识别语言」。下载与应用衔接为一次操作，保存失败仍保留原来启用的语言。为非当前配置应用语言时，会一起启用该配置和所选语言。识别检查跟随界面当前选择。
• 区分 Apple Speech 资源未下载、下载中和暂时无法确认的状态。准备失败后刷新实际资源，并为系统语音服务异常提供对应恢复提示。已有系统资源可能立即可用，不代表在这段时间内下载了完整模型。
• 切换到 Apple 翻译时保留其他翻译服务的已存凭证，识别密钥可独立修改，并保留仅原文字幕和明确同语种时直接显示原文的行为。
• 补齐 Gemini、Azure、xAI、百度、腾讯、火山、DeepL 和 DeepLX 对应模型的语言选项。保留有依据的自动识别，区分中英混合模式，并按来源语言筛选支持的翻译方向。设置、托盘和悬浮窗共用选项与已保存的选择；托盘和悬浮窗也能切换 Gemini 等一体服务的翻译目标，不支持的方向会在保存前拒绝。设置中的识别与目标语言字段保持对齐。
• 明确说明 Mimi 当前接入的 Google Gemini Live 自动识别输入语言，官方列出的 78 个翻译目标均可选择；其他服务仍遵守各自模型的范围。
• 同步 Android 服务目录与协议映射，包括与桌面不同的阿里云实时翻译、识别模型。自定义文字翻译新增的可配置语言不代表用户填写的模型一定支持。
• 桌面端每个配置可显式记住识别与翻译语言。选择或重新应用配置时恢复保存的组合，临时改语言不会覆盖它；未保存组合的配置保留原有行为。Android 暂未提供此编辑器。
• 字幕运行或暂停时可从设置更改语言，并保持暂停状态。修复从仅原文切回 Apple 翻译时误报缺少凭据的问题；编辑配置和准备语言包仍需先停止字幕。
• 在字幕浮窗工具栏和折叠栏增加关闭按钮，点击后停止字幕并隐藏相关窗口。托盘与浮动配置选择器显示服务商图标，避免每次字幕更新都触发设置页无关重绘。
• 托盘与浮动菜单中的语言组合显示在配置名下方，点击字幕顶部的服务信息可直接打开当前配置详情。
• 桌面界面与 README 新增繁体中文、韩语、法语和德语。
• 桌面与 Android 接受火山合法空字幕事件，并累积有界增量预览。#184 报告的原始日语断连仍未复现。
• 收缩字幕条使用简短手势提示，避免遮挡相邻控件，并补齐火山语言错误的本地化恢复入口。
• 新增中英双语服务配置与模型指引。

Apple 翻译仅支持符合条件、运行 macOS 26 或更新版本的 Apple 芯片 Mac，翻译时需要明确选择识别语言。实际可用语言和就绪状态以系统当前结果为准。Intel Mac、Windows、Linux 和 Android 不提供 Apple 翻译。使用云端识别时，音频仍会发送给所选识别服务。

桌面版可在「设置 → 通用」更新；绿色版 ZIP 和 .deb 需手动安装。macOS 沿用固定签名身份，未经 Apple 公证。本次不会为 Android APK 增加 Apple 翻译。

验证：发布准备检查通过：Rust 1,248 项通过（2 项忽略），前端 2,000 项通过，共享核心／JNI、格式、严格 Clippy、lint、类型检查和生产构建均通过。Android Debug／Release 各 145 项通过，包含主机 JNI，两套 lint 通过。统一集成的 macOS、Windows、Linux CI 和 Windows ARM64 编译检查通过。七语浏览器布局与针对性自动回归通过。已签名 macOS UI-only 检查覆盖界面语言切换、当前配置直达、保存组合换行及空白／暂停／折叠字幕控件；这些检查不证明真实采音或服务质量。此前的已签名 macOS 开发版实测通过了英语 → 中文、英语 → 法语的 Apple 文字检查、法语仅原文识别，以及通过系统声音采集的日语 → 中文 Apple Speech + Apple 翻译短合成样本字幕会话；暂停／停止和语言切换也在记录范围内完成检查。资源准备可能复用系统已有资源，首次冷下载及原生资源故障后的同进程恢复仍未验证。语言目录有桌面／Android 共享请求用例覆盖，新增云端语言组合未逐一通过真实音频验证。Windows／Linux 原生行为、Intel Mac 运行、Android 真机、已安装更新和服务持续稳定性仍未验证。测试版本与边界见集成实测记录。

已知限制：百度会话在持续发送音频的静音阶段仍可能报错结束。原因尚未查明，重试不能保证持续恢复。

完整更新记录

mimi v1.5.14

English

• Add Apple Speech local recognition on eligible Apple silicon Macs running macOS 26 or later. Prepare language resources explicitly, then select a ready language. Use original-only subtitles or a separate text translator; remote translation still sends recognized text to that service.
• Switch saved desktop service configurations while listening or paused from Settings, the tray or floating controls. Confirmed subtitles and paused state are preserved; a live switch briefly reconnects. Editing or adding configurations still requires stopping subtitles.
• Keep new live drafts visible after confirmed subtitles, including Gemini sessions after pause/resume or restart.
• Reduce repeated text in continuous Gemini captions, refresh live translation previews sooner, and split long live text at sentence boundaries. Desktop Gemini can prepare a replacement connection when the service announces a planned closure, without restarting capture; a bounded handoff can still omit late output.
• Fix Tencent startup timeouts after a successful handshake on desktop and Android. Improve setup guidance and rejection messages, and align final audio frames and confirmed subtitle pairs across both platforms.
• Show actionable authentication and configuration errors, restore Retry from the floating controls, and keep the saved subtitle background opacity through errors and reconnection outside Immersive Mode.
• Expand provider-specific language choices and preserve automatic source detection for DeepL/DeepLX when a detected language has no explicit local mapping. Available choices remain limited by the selected API and model.
• Give desktop custom recognition profiles an independent display name and an optional declared language list. Service default leaves language selection to the configured recognition service.

Desktop updates are available in Settings → General; portable ZIP and .deb packages require manual installation. Android includes Tencent protocol, Gemini transcript assembly, authentication feedback and language mapping updates. Apple Speech, live saved-service switching, custom recognition profile editing and Gemini's planned connection handoff are desktop features. macOS keeps its existing fixed signing identity and is not Apple-notarized.

Known limitation: Baidu sessions may still fail during silence while audio continues to be sent. The cause remains unresolved, and retry does not guarantee sustained recovery.

Validation: repository checks and Android Debug/Release tests, including host JNI, passed. Short signed macOS sample runs passed for Alibaba, Gemini, Tencent, Volcano, Apple Speech + DeepL and Apple Speech + Index. Windows/Linux native behavior, Intel Mac runtime, Android devices, native Gemini GoAway handoffs and sustained provider stability remain unverified. See the regression record for the tested revisions, failed configurations and limits.

Thanks to @Chtholly000 for their first contribution, improving Gemini continuous captions and planned connection handling in #176.

Full changelog

中文

• 在符合条件、运行 macOS 26 或更新版本的 Apple 芯片 Mac 上新增 Apple Speech 本地识别。手动准备语音资源后，选择已就绪的语言。可只显示原文，也可搭配独立文字翻译；远程翻译仍会将识别文字发送给所选服务。
• 桌面端可在字幕运行或暂停时，从设置、托盘或浮动控制切换已保存的服务配置。保留确认字幕和暂停状态，运行中切换会短暂重连；编辑或添加配置仍需先停止字幕。
• 修复确认字幕之后的新实时草稿不再显示的问题，包括 Gemini 暂停恢复或重启后的字幕。
• 减少 Gemini 连续字幕中的重复文字，让实时译文更快更新，并在句子边界对较长的实时字幕换行。桌面端收到服务计划关闭连接的通知后，可在不重启采音的情况下准备替换连接；切换等待有上限，仍可能遗漏较晚到达的输出。
• 修复桌面端和 Android 的腾讯服务握手成功后仍启动超时的问题。改进配置指引和服务拒绝提示，统一两端的末尾音频帧处理及确认字幕配对。
• 为认证和配置错误提供对应处理入口，修复浮动控制中的重试，并在非沉浸模式的错误和重连期间保持已保存的字幕背景不透明度。
• 扩展各服务对应的语言选项；DeepL／DeepLX 遇到本地没有明确映射的检测语言时，保留服务端自动识别。实际可用语言仍取决于所选 API 和模型。
• 桌面自定义识别配置可单独命名，并可声明语言范围。「服务默认」将语言选择交给配置的识别服务。

桌面版可在「设置 → 通用」更新；绿色版 ZIP 和 .deb 需手动安装。Android 包含腾讯协议、Gemini 字幕组装、认证反馈和语言映射更新。Apple Speech、运行中切换已保存服务、自定义识别配置编辑和 Gemini 计划换连接属于桌面功能。macOS 沿用固定签名身份，未经 Apple 公证。

已知限制：百度会话在持续发送音频的静音阶段仍可能报错结束。原因尚未查明，重试不能保证持续恢复。

验证：仓库检查和 Android Debug／Release 测试已通过，包含主机 JNI。已签名 macOS 的短样本测试通过了阿里云、Gemini、腾讯、火山、Apple Speech + DeepL 和 Apple Speech + Index。Windows／Linux 原生行为、Intel Mac 运行、Android 真机、Gemini 原生 GoAway 换连接及服务持续稳定性仍未验证。测试版本、失败配置和具体边界见回归记录。

感谢 @Chtholly000 的首次贡献，在 #176 中改进 Gemini 连续字幕和计划换连接处理。

完整更新记录

mimi v1.5.13

English

• Choose a subtitle font beside Subtitle display. After selecting it, use Up/Down while the field is focused to switch fonts immediately.
• Configuration and translation-service names save automatically. Repeated edits preserve spaces, punctuation, emoji and the caret; failed saves keep the draft for retry.
• Show recognition and translation services separately when they differ, and one label when shared. Long names stay on one line, with full details on hover.
• Improve saved configuration editing, connection checks before saving, update progress and settings focus outlines. Add Index-Translate setup instructions.

Desktop updates are available in Settings → General; portable ZIP and .deb packages require manual installation. Android has no feature changes in this release. macOS keeps its existing fixed signing identity and is not Apple-notarized.

Validation: repository tests and browser layout checks passed. Signed macOS UI checks cover repeated name edits, font keyboard switching and floating controls. Overlay service combinations and narrow layouts were checked in the browser. Windows/Linux native appearance, Intel Mac runtime, Android devices and real audio/provider quality were not revalidated.

Full changelog

中文

• 字幕显示 右侧可选择字幕字体。选中后保持焦点，按上下方向键即可立即切换。
• 配置名称和翻译服务名称自动保存。连续修改保留空格、标点、emoji 和光标，保存失败可保留草稿重试。
• 识别和翻译来自不同服务时分别显示，同一家合并显示。长名称保持单行，悬停可查看完整信息。
• 改进已保存配置的编辑、保存前连接检查、更新进度和设置焦点边框，补充 Index-Translate 配置说明。

桌面版可在「设置 → 通用」更新；绿色版 ZIP 和 .deb 需手动安装。本次没有 Android 功能变化。macOS 沿用固定签名身份，未经 Apple 公证。

验证：仓库测试和浏览器布局检查已通过；已签名 macOS 界面已验证连续改名、字体键盘切换及浮动控制，服务组合和窄窗布局通过浏览器检查。本轮未重新验证 Windows／Linux 原生外观、Intel Mac 运行、Android 真机和真实音频／服务商质量。

完整更新记录

mimi v1.5.12

English

• Show the current translation service in the expanded subtitle window, using its icon and a short name. The label follows the selected text-translation service, including independent translation profiles; Original mode shows “Original.” Hover or focus to see the profile details, or click to open Settings.
• Keep the service name, latency values and drag handle readable in narrow windows. Updated minimum heights leave room for bilingual subtitles and confirmation times. Immersive and collapsed modes keep the service label hidden.
• Fix Show time for system-audio subtitles. The preference now takes effect immediately for system audio, microphone or both, including Immersive Mode. Times still show when subtitles were confirmed.
• Add Sentence dividers to the floating controls, synchronized with Settings. Immersive Mode keeps the lines hidden while preserving your choice.
• Report a failed Immersive Mode setting save with a brief notification. The previous setting is preserved so you can retry.

Desktop: update in Settings → General or download the package for your platform. Windows portable ZIP and Linux .deb need manual replacement or installation. Versions before v1.3.8 need one manual installation before in-app updates work. macOS retains the fixed self-signed release identity and is not Apple-notarized; use _aarch64.dmg for Apple silicon or _x64.dmg for Intel. On macOS, system-audio capture requires Screen & System Audio Recording access, and microphone capture requires Microphone access. Windows packages are unsigned and may trigger SmartScreen; the portable build requires WebView2. Linux x86_64 remains a preview requiring PulseAudio or PipeWire-Pulse.

Android: download mimi_1.5.12_android.apk (Android 10 or later). This release contains no Android feature changes. Signed v1.5.9–v1.5.11 APKs can update in place. Debug installations use a different signature and must be removed first, deleting their local settings and credentials. The desktop overlay changes above are not Android features. Playback capture remains subject to the source application's restrictions and DRM.

Verification limits: automated checks and browser layout checks cover the changed desktop interface, including Chinese, English and Japanese, narrow windows, translation routes and single/dual audio inputs. Signed macOS UI-only checks verified divider synchronization in both directions between Settings and the floating controls, without provider requests or audio capture. Complete native floating-window visual and cursor checks remain unfinished. Windows/Linux native appearance, Intel Mac runtime, Android devices and real microphone/provider quality were not revalidated for these changes.

Full changelog: v1.5.11...v1.5.12

中文

• 展开的字幕浮窗显示当前翻译服务的图标和短名称，跟随实际选择的文字翻译服务，包括独立翻译配置；原文模式显示「原文」。悬停或键盘聚焦可查看配置详情，点击可打开设置。
• 窄窗口中，服务名称、延迟和拖动柄都有足够空间；同步调整最小高度，避免双语字幕和确认时间被挤压。沉浸和收起模式继续隐藏服务标识。
• 修复系统声音字幕的 显示时间 开关。仅系统声音、仅麦克风及双音源下都会立即生效，沉浸模式也保持一致；时间仍表示字幕确认的时刻。
• 浮窗控制新增 分句线 开关，与设置页同步。沉浸模式继续隐藏分句线，并保留你的选择。
• 沉浸模式设置保存失败时显示简短通知，保留原来的设置，方便重试。

桌面端：通过「设置 → 通用」更新，或下载对应平台安装包。Windows 绿色版 ZIP 和 Linux .deb 需手动替换或安装；早于 v1.3.8 的版本需先手动安装一次，才能使用应用内更新。macOS 沿用固定的自签名发布身份，未经 Apple 公证；Apple 芯片请选择 _aarch64.dmg，Intel 请选择 _x64.dmg。在 macOS 上，采集系统声音需要「屏幕与系统音频录制」权限，采集麦克风需要「麦克风」权限。Windows 安装包未签名，可能触发 SmartScreen；绿色版需要 WebView2。Linux x86_64 仍为预览版，需要 PulseAudio 或 PipeWire-Pulse。

Android：下载 mimi_1.5.12_android.apk（Android 10 或更新版本）。本次没有 Android 功能变化。已签名的 v1.5.9–v1.5.11 正式 APK 可直接覆盖更新；debug 安装签名不同，需先卸载，卸载会删除本机设置和凭据。上述桌面浮窗改动并非 Android 功能。播放声音采集仍受来源应用的捕获限制及 DRM 影响。

验证范围：自动检查和浏览器布局检查覆盖本次桌面界面改动，包括中、英、日文、窄窗口、不同翻译配置及单／双音源。已签名的 macOS UI-only 检查确认分句线在设置页和浮窗之间双向同步，没有调用服务商或采集音频。原生浮窗各状态和光标的完整目视检查仍未完成。本轮没有重新验证 Windows／Linux 原生外观、Intel Mac 运行、Android 设备，以及真实麦克风／服务商识别质量。

完整更新记录： v1.5.11...v1.5.12

mimi v1.5.11

English

• Restore microphone input on desktop. Choose system audio, microphone, or both; the two inputs keep separate subtitles and colors. System audio remains the default, and microphone use requires explicit selection. Switching back to system audio restores the plain subtitle layout while old microphone lines keep a small source icon.
• Bring Show time to subtitle settings and the floating controls. It is off by default and applies while the microphone is selected, including Immersive Mode. Times show when subtitles were confirmed, not the exact start of speech. Source icons and times are easier to read over light backgrounds.
• Add pause/resume to the floating controls and improve the drag handle's tooltip, keyboard access and collapse behavior. Keep confirmed on-screen subtitles when starting another session; clearing subtitles still removes them. This does not enable saved transcript history.
• On Android, scroll through the full current sentence in the expanded panel. Opening the panel locates the current sentence, while reviewing earlier history keeps its position. Action labels stay readable, with two rows on narrow screens or at large system font sizes. Compact and immersive captions retain their existing line limits.

Desktop: update in Settings → General or download the package for your platform. Windows portable ZIP and Linux .deb need manual replacement or installation. Versions before v1.3.8 need one manual installation before in-app updates work. macOS retains the fixed self-signed release identity and is not Apple-notarized; use _aarch64.dmg for Apple silicon or _x64.dmg for Intel. On macOS, system-audio capture requires Screen & System Audio Recording access, and microphone capture requires Microphone access. Windows packages are unsigned and may trigger SmartScreen; the portable build requires WebView2. Linux x86_64 remains a preview requiring PulseAudio or PipeWire-Pulse.

Android: download mimi_1.5.11_android.apk (Android 10 or later). Signed v1.5.9 and v1.5.10 APKs can update in place. Debug installations use a different signature and must be removed first, deleting their local settings and credentials. Desktop microphone selection, source colors and time controls are not Android features. Playback capture remains subject to the source application's restrictions and DRM.

Verification limits: signed macOS UI-only checks and Android 15 emulator UI checks used synthetic subtitles without provider requests or audio capture. Windows/Linux native appearance, Intel Mac runtime, physical Android devices and real microphone/provider quality were not revalidated. Automated tests and package verification do not replace those checks. Platform/state screenshots and original image archive are retained in the PR.

Full changelog: v1.5.10...v1.5.11

中文

• 桌面端恢复麦克风输入，可以选择系统声音、麦克风，或同时开启两路；两路字幕和颜色分别保留。默认仍只采集系统声音，麦克风需要主动选择。切回仅系统声音后恢复原来的字幕样式，之前的麦克风字幕保留一个小来源图标。
• 字幕设置和悬浮控制里补回 显示时间 开关。默认关闭，开启麦克风后生效，沉浸模式也跟随这个开关。时间表示字幕确认的时刻，不是精确的说话起点。来源图标和时间在亮色背景下更容易看清。
• 悬浮控制新增暂停／继续，并改进拖动柄的提示、键盘操作和收起行为。开始下一次会话时保留当前窗口内已确认的字幕，点击清空仍可移除；这不会自动开启字幕历史保存。
• Android 展开面板可以滚动读完当前长句。打开时定位当前句，回看前面的历史时保留阅读位置。操作按钮不再挤成多行文字，窄屏或大系统字号下会分成两排。紧凑和沉浸字幕仍保留原有行数限制。

桌面端：通过「设置 → 通用」更新，或下载对应平台安装包。Windows 绿色版 ZIP 和 Linux .deb 需手动替换或安装；早于 v1.3.8 的版本需先手动安装一次，才能使用应用内更新。macOS 沿用固定的自签名发布身份，未经 Apple 公证；Apple 芯片请选择 _aarch64.dmg，Intel 请选择 _x64.dmg。在 macOS 上，采集系统声音需要「屏幕与系统音频录制」权限，采集麦克风需要「麦克风」权限。Windows 安装包未签名，可能触发 SmartScreen；绿色版需要 WebView2。Linux x86_64 仍为预览版，需要 PulseAudio 或 PipeWire-Pulse。

Android：下载 mimi_1.5.11_android.apk（Android 10 或更新版本）。已签名的 v1.5.9、v1.5.10 正式 APK 可直接覆盖更新；debug 安装签名不同，需先卸载，卸载会删除本机设置和凭据。桌面端的麦克风选择、来源颜色与时间开关并非 Android 功能。播放声音采集仍受来源应用的捕获限制及 DRM 影响。

验证范围：已签名的 macOS UI-only 检查和 Android 15 模拟器界面检查均使用合成字幕，没有调用服务商或采集音频。本轮没有重新验证 Windows／Linux 原生外观、Intel Mac 运行、Android 真机，以及真实麦克风／服务商识别质量；自动测试和打包验证不能替代这些检查。各平台、各状态截图及原图包 已保留在 PR。

完整更新记录： v1.5.10...v1.5.11

mimi v1.5.10

English

• Add Show subtitles sooner to desktop subtitle settings, the menu-bar panel and the floating panel. It is on by default: subtitles appear early and may keep changing. Turn it off to wait for confirmed text; continuous speech can mean a longer wait. You can change it during a session. This changes display timing, not recognition or translation accuracy.
• Remove Keep text opaque and restore the normal subtitle appearance.
• Show application icons in the macOS audio picker and fix typing in the floating panel's application search.
• Save desktop API keys in Mimi's private local files on macOS, Windows and Linux. Existing keys move automatically, with no migration screen or completion notice. Mimi verifies the saved copy before removing the old system-store entry. After migration, switching configurations, restarting and updating use only the local files. Importing old keys may still require the system store's authorization. Keys are stored as plaintext protected by local file permissions.
• On macOS, opening the application audio picker no longer requests recording access. If recording permission is denied or the user stops capture through macOS, Mimi stops retrying automatically. Development and release builds keep separate signing settings to avoid unnecessary permission changes. Setup & help explains how to remove and re-add a stale recording permission entry.

Desktop: update in Settings → General or download the package for your platform. Windows portable ZIP and Linux .deb need a manual replacement or installation. Versions before v1.3.8 need one manual installation before in-app updates work. macOS retains the fixed self-signed identity and is not Apple-notarized; use _aarch64.dmg for Apple silicon or _x64.dmg for Intel. Starting system-audio capture still requires macOS recording permission. Windows packages are unsigned and may trigger SmartScreen; the portable build requires WebView2. Linux x86_64 remains a preview requiring PulseAudio or PipeWire-Pulse; Secret Service is used only to import old credentials.

Android: download mimi_1.5.10_android.apk (Android 10 or later). The signed v1.5.9 APK can update in place. Debug installations use a different signature and must be removed first, deleting their local settings and credentials. The new subtitle controls and local credential migration described above are desktop changes. Android playback capture remains subject to the source application's restrictions and DRM.

Verification limits: native credential migration and configuration switching were checked on macOS. Windows/Linux native credential migration, Intel Mac runtime and physical Android behavior remain unverified. Automated checks and package builds do not replace those device tests. Recognition and translation quality still depend on the selected provider and model.

Full changelog: v1.5.9...v1.5.10

中文

• 桌面字幕设置、菜单栏面板和浮窗面板新增 更早显示字幕。默认开启，字幕会提前出现，文字可能继续修改；关闭后等字幕确认再显示，连续讲话时可能等得更久。播放中也可以切换。它只改变显示时机，不提高识别或翻译准确率。
• 移除 保持文字不透明，恢复正常的字幕显示效果。
• macOS 音频应用选择器显示应用图标，并修复浮窗中应用搜索框无法输入的问题。
• macOS、Windows 和 Linux 的 API key 改为保存在 Mimi 的本机私有文件中。旧密钥自动迁移，不增加迁移页面或完成通知；确认新副本保存成功后，再删除系统存储中的旧条目。迁移后，切换配置、重启和更新都只使用本地文件。首次读取旧密钥时，系统仍可能要求授权。密钥以明文保存，由本地文件权限保护。
• macOS 打开音频应用选择器时不再请求录音权限；权限被拒绝或用户通过 macOS 停止采集后，不再自动反复重试。开发版和正式版分别保留自己的签名设置，避免无意换签导致重新授权。使用与常见问题 补充了删除旧权限条目、重新添加应用的恢复方法。

桌面端：通过「设置 → 通用」更新，或下载对应平台安装包。Windows 绿色版 ZIP 和 Linux .deb 需手动替换或安装；早于 v1.3.8 的版本需先手动安装一次，才能使用应用内更新。macOS 沿用固定自签名身份，未经 Apple 公证；Apple 芯片请选择 _aarch64.dmg，Intel 请选择 _x64.dmg。启动系统声音采集仍需 macOS 录音权限。Windows 安装包未签名，可能触发 SmartScreen；绿色版需要 WebView2。Linux x86_64 仍为预览版，需要 PulseAudio 或 PipeWire-Pulse；Secret Service 仅用于迁移旧凭据。

Android：下载 mimi_1.5.10_android.apk（Android 10 或更新版本）。已签名的 v1.5.9 正式版可直接覆盖更新；debug 安装签名不同，需先卸载，卸载会删除本机设置和凭据。上面的新字幕开关与本地密钥迁移属于桌面端改动。Android 播放声音采集仍受来源应用的捕获限制及 DRM 影响。

验证范围：已在 macOS 检查真实凭据迁移和配置切换。Windows／Linux 原生凭据迁移、Intel Mac 运行及 Android 真机行为仍未验证，自动检查和成功打包不能替代这些设备测试。识别和翻译质量仍取决于所选服务商及模型。

完整更新记录： v1.5.9...v1.5.10

mimi v1.5.9

English

• Correct the macOS ScreenCaptureKit sample-rate setter to use the native integer type, with configuration regressions for 16 and 24 kHz.
• Preserve distinct repeated final subtitles when Stop flushes a sentence while an earlier identical sentence is still being translated. Duplicate confirmations from the same source sentence remain idempotent.
• Restore Gemini Live connection setup on desktop and Android, accept language-only transcription updates, and keep continuous source/translation blocks together. Confirmation uses a stability checkpoint because this model does not consistently provide utterance boundaries.
• Keep multilingual settings readable in narrow windows, improve contextual help and labeled saved-value actions, and use short-lived settings notifications without moving page content. Proxy route choices save immediately; custom addresses save on Enter or when leaving the address group. Diagnostics distinguish interface round trips from translation-request time.
• Fix Linux overlay click-through with empty native input regions and preserve the latest lock/unlock state across window lifecycle changes.
• Share subtitle assembly and translation policy between desktop and Android, with shared contracts and Android JNI coverage. Preserve repeatable private audio test inputs, failed cases and evidence boundaries for future regression work.

Recognition and translation quality still depend on the selected provider and model. These fixes do not establish a general accuracy improvement. Gemini's continuous-block checkpoint is a bounded heuristic, not a guaranteed service utterance boundary. Long-subtitle visibility, physical Android behavior, Intel Mac runtime, and real Windows/Linux provider sessions remain separate acceptance work.

Desktop: update in Settings → General or download the package for your platform. Windows portable ZIP and Linux .deb require a manual replacement or installation. Versions before v1.3.8 need a manual installation before in-app updates work. macOS keeps the fixed self-signed identity and is not Apple-notarized; choose _aarch64.dmg for Apple silicon or _x64.dmg for Intel. Windows packages are unsigned and may trigger SmartScreen; the portable build requires WebView2. Linux x86_64 remains a preview requiring PulseAudio or PipeWire-Pulse and Secret Service.

Android: download mimi_1.5.9_android.apk (Android 10 or later). The signed v1.5.8 release can update in place; debug installations use a different signature and must be removed first, deleting their settings and saved credentials. Android captures system playback only; applications that prohibit playback capture and DRM-protected audio cannot be captured. Desktop presentation changes do not imply matching Android features.

Verification scope: signed macOS development tests observed nonzero system audio and complete Alibaba recognition/translation/history chains from two official Japanese trailer extracts; the Conan subtitle window was inspected directly. No independent dialogue or bilingual accuracy score was established. Automated checks, native visibility, provider behavior and package verification are recorded separately; a successful package build does not establish physical-device acceptance. Production credentials remain in the OS secure store, and release packages exclude the development debugger and private evidence recorder.

Full changelog: v1.5.8...v1.5.9

中文

• macOS ScreenCaptureKit 采样率设置改为原生整数类型，补齐 16 和 24 kHz 配置回归。
• 停止时，若较早的相同句子仍在翻译，保留新出现的重复尾句；同一句来源的重复确认仍不会重复接纳。
• 恢复桌面和 Android 的 Gemini Live 连接配置，接纳仅含语言的转录更新，并让连续原文和译文保持整块配对。由于该模型不稳定提供语句边界，完整块确认使用稳定检查点。
• 改善窄窗口中的多语言设置布局、上下文帮助和带文字的已保存值操作；短暂通知不再推动页面内容。代理路线选择立即保存，自定义地址在按 Enter 或离开地址组时保存；诊断明确区分接口往返与翻译请求耗时。
• Linux 浮窗穿透改用真正的原生空输入区域，并在窗口生命周期变化中保留最新锁定／解锁状态。
• 桌面和 Android 共用字幕组装及翻译策略，补齐共享契约和 Android JNI 检查；保留可重放的私有语音输入、旧失败案例及证据边界，供后续回归验证使用。

识别和翻译质量仍取决于服务商及模型，本次修复尚未证明普遍的准确度提升。Gemini 的连续整块确认使用有界启发式，不保证对应服务语句边界。长字幕完整可见性、Android 真机行为、Intel Mac 运行，以及 Windows／Linux 的真实服务会话仍需分别验收。

桌面端：通过「设置 → 通用」更新，或下载对应平台安装包。Windows 绿色版 ZIP 和 Linux .deb 需手动替换或安装；早于 v1.3.8 的版本需先手动安装一次，才能使用应用内更新。macOS 沿用固定自签名身份，未经 Apple 公证；Apple 芯片请选择 _aarch64.dmg，Intel 请选择 _x64.dmg。Windows 安装包未签名，可能触发 SmartScreen；绿色版需要 WebView2。Linux x86_64 仍为预览版，需要 PulseAudio 或 PipeWire-Pulse 及 Secret Service。

Android：下载 mimi_1.5.9_android.apk（Android 10 或更新版本）。已签名的 v1.5.8 正式版可直接覆盖更新；debug 安装签名不同，需先卸载，卸载会删除其设置和已保存的凭据。Android 仅采集系统播放声音，无法采集禁止播放捕获的应用及受 DRM 保护的音频；桌面显示改动不代表 Android 有对应功能。

验证范围：签名 macOS 开发版使用两段官方日语预告提取音频，观察到非零系统声音与完整的阿里云识别／翻译／历史链；柯南实际字幕窗口已直接检查。尚未建立独立对白或双语准确率评分。自动检查、原生可见性、服务行为和安装包验证分别记录；成功构建安装包不代表真实设备验收。正式版凭据继续保存在系统安全存储中，发布包不包含开发调试器和私有取证录制功能。

完整更新记录： v1.5.8...v1.5.9

mimi v1.5.8

English

• Add system-audio capture from a selected application on macOS and Windows build 20348 or later, alongside the default all-applications capture. Settings and the floating controls share a searchable application picker. Changing the target preserves confirmed subtitles and paused state; a missing target reports an error without silently capturing every application. Browser tabs remain part of their browser application; Linux and Android retain their existing playback capture.
• Preserve complete macOS audio buffers when data spans native memory blocks and correct planar multichannel decoding. These fixes address capture-byte and channel-layout loss risks without claiming a general recognition-accuracy improvement.
• Restore Live Subtitles and Immersive Mode controls in Subtitle settings, including resume and action-error feedback. Add Skip translation to supported profiles and floating controls for recognition-only output, and keep hand cursors consistent across interactive desktop controls.
• Save independent recognition and text-translation proxy choices with each desktop service profile; integrated services use one route. Add explicit paste actions for configuration fields and scoped reveal controls for saved keys, and improve active-service status feedback.
• Add a subtitle background transparency slider, defaulting to 20% transparency, with the saved value restored after Immersive Mode. Keep older history readable, offer Keep text opaque in the floating controls, and place the custom color picker directly after the preset swatches. These presentation controls are desktop features.
• Add distinct ChatMock and OpenAI-compatible text-translation choices on desktop and Android, with independent saved destinations and credentials. Android now supports DeepL, DeepLX, ChatMock, OpenAI-compatible translation or original-only output after Alibaba realtime recognition. Filter complete leading ChatMock reasoning blocks and reject unfinished translation responses; shared protocol fixtures maintain the desktop/Android contract. ChatMock must be run and authenticated separately; localhost refers to the device running Mimi.
• Bundle the GLES entry library required by the Linux AppImage and recover an explicit first credential save when Secret Service has no default collection. Explicit Save can retry failed credential reads after storage is restored; native authorization remains required.
• Preserve complete original/translation preview pairs for custom DashScope and OpenAI speech recognition. A later partial translation prefix no longer replaces an already completed preview.
• Correct the overlay minimum height for independent subtitle tracks, reserving space for each source's current subtitle block at the selected font size.
• Wait for the final subtitle publication attempt before stopping finishes, including an accepted trailing final translation. This confirms publication was attempted; subscriber receipt and native painting remain separate steps.
• Temporarily hide desktop microphone selection, microphone subtitle color settings and microphone setup guidance. System audio remains the default. On startup, a saved microphone or both-input selection changes to system audio and its recording opt-in is cleared; enable recording again explicitly if needed. Saved sessions remain available.
• Set new-installation overlay defaults to 659×328 and font size 16. Existing valid window size and font preferences are preserved.

Recognition accuracy is still limited by the selected provider and model. Real-time Japanese and French recognition can expand or repeat text; this release does not establish a general accuracy fix. The subtitle boundary fixes leave recognition parameters and translation prompts unchanged. Production credentials remain in the OS secure store.

Desktop: update in Settings → General or download the package for your platform. Windows portable ZIP and Linux .deb require a manual replacement or installation. Versions before v1.3.8 need a manual installation before in-app updates work. macOS keeps the fixed self-signed identity and is not Apple-notarized; choose _aarch64.dmg for Apple silicon or _x64.dmg for Intel. Windows packages are unsigned and may trigger SmartScreen; the portable build requires WebView2. Linux x86_64 remains a preview requiring PulseAudio or PipeWire-Pulse and Secret Service.

Android: download mimi_1.5.8_android.apk (Android 10 or later). The independent text-translation features above are included; the new desktop capture picker and subtitle presentation controls do not imply matching Android features. Android captures system playback only; applications that prohibit playback capture and DRM-protected audio cannot be captured. The signed 1.5.7 release can update in place; debug installations use a different signature and must be removed first, deleting their settings and saved credentials.

Verification: the local canonical desktop check passed 1,045 Rust tests and 1,075 frontend tests, formatting, strict Clippy, ESLint, TypeScript and the production build. Signed macOS UI-only inspection confirmed font size 16, hidden microphone input/color controls and a working system application picker. UI-only inspection does not establish live capture or provider behavior. Windows/Linux provider sessions, Intel Mac runtime, live captured-audio translation through ChatMock and physical Android behavior remain unverified for this release. Available languages and third-party compatibility depend on the selected services and models.

Full changelog: v1.5.7...v1.5.8

中文

• macOS 和 Windows build 20348 或更新版本新增指定应用的系统音频采集，默认仍采集所有应用。设置与浮窗控制面板共用可搜索的应用选择器；切换目标保留已确认字幕和暂停状态，目标不可用时报告错误，不会悄悄改为采集所有应用。浏览器标签页仍属于同一个浏览器应用；Linux 和 Android 保留原有播放声音采集方式。
• macOS 音频数据跨越原生内存块时读取完整缓冲，并修正平面多声道解码。这些修复处理采集字节和声道布局的丢失风险，不代表普遍的识别准确度提升。
• 在字幕设置中恢复「实时字幕」和「沉浸模式」控制，包含继续运行及操作失败反馈；支持仅识别的服务配置与浮窗控制面板新增「跳过翻译」，桌面可交互控件保持一致的手形光标。
• 桌面每套服务配置分别保存识别与文字翻译的代理选择，集成服务使用单一路由；配置字段新增明确的粘贴操作，已保存密钥可按当前配置查看，并改进当前服务的状态反馈。
• 新增字幕背景透明度滑块，默认透明度为 20%，退出沉浸模式后恢复已保存的值；保持旧字幕记录可读，在浮窗提供「保持文字不透明」，自定义颜色选择器紧接预设色块显示。这些显示控制属于桌面功能。
• 桌面和 Android 将 ChatMock 与 OpenAI 兼容文字翻译分为独立选项，分别保存服务地址和凭据。Android 可在阿里云实时识别后选择 DeepL、DeepLX、ChatMock、OpenAI 兼容翻译或仅原文；过滤完整的 ChatMock 前置思考块，拒绝未完成的翻译响应，共享协议用例维护两端契约。ChatMock 需自行运行并完成账号认证；localhost 指运行 Mimi 的设备。
• Linux AppImage 补入所需的 GLES 入口库；Secret Service 尚无默认集合时，可恢复用户明确触发的首次凭据保存。恢复安全存储后再次点击保存，可重试失败的凭据读取；系统授权仍需正常完成。
• 自定义 DashScope 和 OpenAI 语音识别保留完整配对的原文与译文预览；已经完成的预览不再被后续较短的翻译片段替换。
• 修正独立字幕轨道的浮窗最低高度，按所选字号为每路来源当前的字幕块留出空间。
• 停止完成前等待最后一次字幕发布尝试，包含已接受的尾句最终译文。这保证已尝试发布；订阅端收到事件和原生界面绘制仍是后续步骤。
• 暂时隐藏桌面麦克风选择、麦克风字幕颜色设置和麦克风使用引导。系统音频仍为默认输入；启动时会把旧的麦克风或双路输入设置改为系统音频，并清除录音开关，需要录制时请重新明确开启。已保存的记录仍可使用。
• 新安装的浮窗默认尺寸改为 659×328、字号改为 16；已有的有效窗口尺寸和字号设置会保留。

识别准确度仍取决于所选服务商和模型。日语、法语实时识别仍可能扩写或重复文字，本版本尚未证明普遍的准确度修复。字幕边界修复未改变识别参数和翻译提示词；正式版凭据继续保存在系统安全存储中。

桌面端：通过「设置 → 通用」更新，或下载对应平台安装包。Windows 绿色版 ZIP 和 Linux .deb 需手动替换或安装；早于 v1.3.8 的版本需先手动安装一次，才能使用应用内更新。macOS 沿用固定自签名身份，未经 Apple 公证；Apple 芯片请选择 _aarch64.dmg，Intel 请选择 _x64.dmg。Windows 安装包未签名，可能触发 SmartScreen；绿色版需要 WebView2。Linux x86_64 仍为预览版，需要 PulseAudio 或 PipeWire-Pulse 及 Secret Service。

Android：下载 mimi_1.5.8_android.apk（Android 10 或更新版本）。包含上述独立文字翻译功能；新增桌面采集选择器和字幕显示控制不代表 Android 有对应功能。Android 仅采集系统播放声音，无法采集禁止播放捕获的应用及受 DRM 保护的音频。已签名的 1.5.7 正式版可直接覆盖更新；debug 安装签名不同，需先卸载，卸载会删除其设置和已保存的凭据。

验证：本地桌面标准检查通过 1,045 个 Rust 测试、1,075 个前端测试，同时通过格式检查、严格 Clippy、ESLint、TypeScript 和生产构建。签名 macOS 原生 UI-only 检查确认字号 16、麦克风输入及颜色入口隐藏，系统应用选择器正常；UI-only 检查不代表真实采集或服务商行为已验证。本版本的 Windows/Linux 服务会话、Intel Mac 运行、通过 ChatMock 翻译真实采集音频及 Android 真机行为仍未验证。可用语言和第三方兼容性取决于所选服务与模型。

完整更新记录： v1.5.7...v1.5.8