Glimpse

English — Scroll down for the English README.

Glimpse

关键词：动态高亮、选择高亮、持久高亮、正则查询、捕获组、自定义 CSS、CSS 自动补全、高亮器描述、高亮索引、提词器、光标联动、滚动条标记、缩略图、文字渐变、字幕投影、滚动同步、穿透锁定、文档绑定、分组管理、导入导出、正则查找替换、替换预览、音乐播放、歌词面板、卡拉OK逐字高亮、在线歌词、多平台下载、笔记朗读、语音合成（TTS）、播放条、内容过滤、标点细分高亮、音频导出、生成音频文件、GPT-SoVITS、Qwen3-TTS、Windows 本地语音

根据选中内容或搜索关键词动态高亮文本的 Obsidian 插件，主要功能：

• 选择高亮：选中文本后即时高亮全文所有匹配（仅在检索到选区之外的至少一个出现位置时才出现装饰），附滚动条标记与缩略图
• 持久高亮：按正则/关键词查询持久标记，支持捕获组、父行、开始/结束 widget、自定义 CSS（内置带 CSS 自动补全的代码编辑器）与高亮器描述，配标签组管理与一键导入导出
• 高亮索引：自动检索 ==高亮== 文本，按文档标题层级组织为侧边栏索引，支持光标联动与键盘导航
• 提词器（桌面端）：歌词式浮动提词窗口，跟随文档/光标实时显示，支持 100–900 字重、文字渐变、字幕投影、穿透锁定、文档绑定、滚动同步与多实例
• 朗读（桌面端）：把当前笔记念出来——先按 17 项规则把 Markdown 过滤成「该被听见的内容」，再按标点三级分段、逐段合成、边播边取下一段，同时高亮正在读的那一块（可细到当前分句）；可从文档开头 / 光标处 / 选区起读，支持预取、暂停续播、自动滚动、光标跟随与三种 TTS 引擎（本地 GPT-SoVITS / Qwen3-TTS / 零安装的 Windows 本地语音）；选中文字还可一键导出为单个音频文件
• 正则查找替换：当前笔记（或选区）内的正则查找、替换与替换预览——Ctrl+H / 命令 / 右键呼出，命中的文本实时点亮在编辑器里，替换前后对照与匹配列表可点击定位，支持 $1 / $<name> 捕获组引用与 \n 转义，逐处写回、一步撤销（移植自 obsidian-regex-replace，详见致谢）
• 音乐：侧边栏歌词音乐面板——三标签页歌单（网易云账号歌单同步到本地 / 在线歌单四平台搜索下载 / 本地歌单自定义分组与拖拽排序）、逐字卡拉OK高亮与双语注释、歌词自动跟随（手动滚动即暂停）、音频文件夹（MP3/FLAC/M4A/OGG，支持库外绝对路径）作歌单、四平台在线歌词获取、下载前试听与推荐歌单

目前仅支持源码模式（Source）和实时预览模式（Live Preview）。阅读模式（Reading）和旧版编辑器暂不支持。

选择高亮

有选中内容时，高亮选中文本的所有出现位置：

• 大小写不敏感匹配，全文档检索（不限于可见区域）
• 装饰出现条件：仅当检索到选区之外的至少一个出现位置时才施加装饰；选中仅出现一次的文本时不显示任何下划线，滚动条标记同样不出现
• 当前选中文本标记为 .cm-current-string
• 文档中其他位置匹配的字符串标记为 .cm-matched-string
• 所有匹配项附带 data-contents 属性存储当前选中字符串值
• 滚动条标记 / 缩略图：检索到结果时滚动条显示匹配位置标记；可选在编辑器右侧显示缩略图（类似 VS Code minimap，可拖动滚动）
• 检索上限：设置中可调「选择检索的字符串上限」（2-60，默认 30），超过该长度的选中文本不再进行全文匹配，避免超长选择拖慢编辑

持久高亮

定义搜索查询并关联 CSS 类名和颜色来创建持久高亮。匹配的字符串会自动标记对应 CSS 类，并按所设背景色着色（背景色可留空，未选色时只应用高亮类，样式完全交给自定义 CSS）。每条样式与颜色缓存于索引，标记可随时开关控制匹配表现。

每条样式可附带一段描述，备注该表达式匹配什么内容——仅用于设置页展示，不参与匹配。设置页里描述排在名称下一行，列表卡片里排在表达式下一行。

支持正则表达式查询（需开启对应选项），正则模式下可使用命名捕获组（如 (?<groupName>…)）或无名捕获组（如 (…)）精确高亮子匹配内容。

标记类型

每条样式可组合启用多种标记方式：

• 匹配：高亮完整匹配文本（默认开启）
• 父行：为匹配所在整行添加 CSS 类，支持针对整行而非单词设置样式
• 开始 / 结束：在匹配起止位置插入零宽度 widget 元素，配合 CSS 可实现前缀/后缀图标
• 捕获组：正则模式下，高亮捕获组匹配的子内容而非完整表达式。子内容使用自定义样式颜色。支持命名组 (?<name>…) 和无名组 (…)，勾选「捕获组」时自动跳过整段匹配

自定义 CSS

每条样式可编写独立 CSS 规则，经 CSSStyleSheet + document.adoptedStyleSheets 注入页面（不创建 <style> 元素）。
编辑器内通过 CodeMirror 实例高亮渲染，支持深色/浅色主题适配。CSS 变更随保存即时生效。

输入框本身是一个完整的 CodeMirror 6 代码编辑器：语法高亮、括号与引号自动闭合、撤销历史，并按输入自动补全 CSS——属性名、属性值关键字与选择器标签名都会在输入时弹出候选，也可用 Ctrl/Cmd-Space 手动唤起；选中属性会连带插入冒号与空格。

标签组

样式支持分组管理：新建、重命名、删除分组，拖拽样式至标签页即可归类，分组标签可拖动调整顺序（默认分组固定首位不可拖动）。「全部启用/禁止」工具栏按钮统一控制当前分组匹配开关状态。

导入导出

支持一键导入（从剪贴板 JSON 或文件）、一键导出（批量导出所有样式含分组信息）、单条样式导出。导入/导出弹窗使用样式化纯文本框：按钮置于文本框上方（导出到文件 / 导出到粘贴板、从文件导入 / 从下面粘贴板导入），文本框高度随内容自适应，界面风格统一。兼容旧格式数据（无分组 → 归入"默认"）。

Pandoc 导出中的自定义样式高亮（:::）

{
  "Pandoc-Syntax": {
    "class": "Pandoc-Syntax",
    "color": "#77787C4A",
    "regex": true,
    "query": "::: \\{.*?\\}[\\s\\S]*?:::",
    "mark": ["match", "group"],
    "css": ""
  }
}

高亮索引

自动检索当前文档中 ==...== 包裹的 Obsidian 标准高亮文本，以标题层级组织在侧边栏展示。

• 标题分组：按文档标题结构组织高亮项，无标题的高亮项归入文档名分组展示
• 层级图标：h1-h6 标题前显示对应 lucide heading 图标，颜色跟随 --h1-color ~ --h6-color CSS 变量
• 祖先标题栏：共享祖先标题以独立行展示于首个索引项前，与正常标题使用相同图标和配色
• 毛玻璃卡片：高亮文本以圆角毛玻璃卡片展示，支持悬浮增强毛玻璃特效
• 光标联动：编辑器光标落入高亮语句所在行时，对应卡片自动选中并滚动至屏幕中央
• 键盘导航：聚焦索引标签页时 ↑/↓ 切换上/下一项高亮卡片（从已选中卡片起算），并同步编辑器光标与提词器
• 右键复制：右键卡片复制文本内容并弹出提示
• 锚定文档回退：当前页面检索不到高亮时，自动从提词器锚定（绑定）的文档中检索展示；锚定文档未打开为视图时读盘检索
• 自动刷新：首次打开文档即检索，切换文档时自动刷新索引（补监听 file-open，覆盖新建 / 资源管理器点击当前标签页等不触发叶子变更的打开方式）；新焦点文档无高亮时保留上次检索结果
• 命令面板：注册「打开高亮索引」命令，可从命令面板呼出
• 设置开关：「启动时默认打开高亮索引」（默认关闭），开启后插件启动时自动启用索引标签页

提词器（桌面端）

桌面歌词式浮动提词窗口，跟随文档内容实时显示，适合提词、朗读、讲解场景。仅桌面端可用。

• 四种内容模式：
  • 行提取：默认静态显示（打开/切换文档时提取一次），点击工具栏「跟踪光标」开启光标跟随——光标移动自动提取所在行（轮询检测，绑定/活动文档均可）；滚轮/按钮可手动切换上一行、下一行
  • 高亮提取：按顺序显示 ==...== 高亮匹配，适合按稿朗读
  • 歌词提取：展示音乐模块当前正在播放的歌词行，随播放进度自动换行（按歌曲 + 行号去重，不被高频状态推送重绘）；上一项/下一项/滚轮跳转到相邻歌词行的时间戳（暂停时同样生效）；无播放会话时显示「未在播放歌曲」占位，前奏未唱到第一行显示「♪ 前奏 ♪」
  • 朗读提取：展示朗读当前正在读的那一块，随播放自动推进（开启「高亮按标点细分」时细到当前分句）；内容直接取编辑器里的原文片段，因此 Markdown 记号会正常渲染；未在朗读时显示「未在朗读」占位。内容由音频时钟驱动，没有可 seek 的「上一项/下一项」
  • 选中提取：选中文本时临时覆盖显示选中内容，取消选中自动恢复
• 模式切换：工具栏自绘下拉（逐行提取 / 高亮提取 / 歌词提取 / 朗读提取）——带 chevron-down 下拉箭头，弹出菜单跟随 Obsidian 主题配色并高亮当前项（原生 <select> 的弹层是系统级窗口，既不随顶栏显隐也不跟随主题，故未采用），收起值与弹出选项均居中；点击提词器之外、按 Esc、鼠标离开窗口都会收起；窗口拖拽/滚轮逻辑对下拉整块豁免（选项区可正常滚动），穿透锁定时隐藏
• 高亮索引联动：点击高亮索引卡片自动绑定该文档并切换到高亮提取模式；双击提词器文本区同步选中索引中对应卡片；开启滚动同步后，上一项/下一项切换同样联动选中对应卡片
• 滚动同步：工具栏「穿透锁定」右侧按钮（lucide link）开启后，上一项/下一项切换同步触发——逐行模式光标跳转对应上一/下一行，高亮模式选中索引中对应上一/下一项卡片；穿透锁定时随非交互按钮一并隐藏
• 文档绑定：将窗口固定到某个文档，不再跟随活动文档；已锁定后再次点击绑定按钮直接解除锁定，不会转向锁定当前活动文档
• 穿透锁定：窗口整体穿透鼠标（仅保留交互按钮），不遮挡编辑；背景全透明统一由「隐藏背景」按钮控制，穿透锁定本身不再改动背景；锁定状态下按钮栏同样仅在鼠标悬停时显示，移出窗口自动隐藏，「跟踪光标」「滚动同步」按钮随其他非交互按钮一并隐藏
• 隐藏背景：激活后窗口背景全透明（悬停/拖拽也不显示），未激活时背景色常显且透明度引用设置界面所设的背景透明度；激活时正文文字自动附加字幕投影
• 文字投影（字幕效果）：隐藏背景时为正文文字附加柔和投影（drop-shadow，严格位于文字之下），设置「文字阴影」折叠分组可调投影开关、水平/垂直偏移、模糊半径、不透明度（均带重置按钮），默认右偏 2px/下偏 2px/模糊 2px/不透明度 40%
• 文字描边：正文文字轮廓描线（桌面歌词风格，复杂背景上更易读，始终生效），默认开启、外侧描边、1px 白色；设置「文字描边」折叠分组可调描边开关、描边方式、线宽（0-12px）与描边颜色（色板，「清除」恢复默认白），带重置按钮。描边方式二选一：字形描边（-webkit-text-stroke + paint-order: stroke fill，描边垫在填充之下、可见线宽约为一半，最锐利）与外侧描边（SVG feMorphology 膨胀滤镜，作用于渲染后像素、渐变/逐字渐变均兼容，完整线宽外露；滤镜定义由插件注入文档并随设置同步，大线宽下斜角略方）。两种方式均不增粗笔画，可与隐藏背景时的投影叠加
• 文字渐变：正文文字渐变色，设置「文字渐变」折叠分组提供渐变开关（覆盖「字体颜色」）、线性/径向类型、整体/逐字范围（逐字模式每字独立裁切渐变，emoji 等组合字符不拆碎）、渐变角度与颜色停靠点编辑器（增删/位置/色板，按位置排序），附与提词器同参数的实时预览
• 宽度自适应：背景贴合内容 —— 可见面板以 CSS fit-content 直接包住正文（即使宽度状态因异步异常失配，可见背景也不留大片空白），根窗口宽度按内容最宽行与字号实时计算并钳制视口上限，长文本换行不溢出；工具栏不参与宽度计算，窗口比工具栏窄时悬停工具栏自动换行收纳；适配多重触发（渲染完成 + ResizeObserver 捕捉字体迟载/主题切换等事后变化 + 视口 resize + 每秒兜底再同步），宽度失配自动收敛；根窗口自身不拦截鼠标（失配的透明区域不会挡住背后编辑器），拖宽把手始终贴着可见面板右缘；宽度锁定或拖宽中背景铺满所选宽度（拖到哪面板就多宽），右缘拖拽后自动锁定，锁定后仍可拖拽，新宽度继承为锁定宽度
• 拖拽吸附：贴近视口边缘或中心线时自动吸附，附辅助线提示
• 位置稳定：窗口位置由用户决定，内容高度变化（换行、改字号、宽度自适应重算）不会移动窗口；重启后按保存的位置原样恢复，不再因恢复后的行文变高被抬升
• 字体大小：32/40/50/64/80px 五档循环
• 字体 / 字重 / 颜色：设置中可调正文字体（本机字体选择模态窗——取消/确定常驻标题行、列表再长也无需滚动，queryLocalFonts() 枚举、兜底候选表测宽；多选 + 拖拽调优先级，首个可用字体优先生效、缺失自动顺延；支持搜索、预览与自定义字体输入）、字重（跟随主题 / 100–900）与文字颜色（色板，点击即应用；「清除」恢复跟随主题），均带「重置为初始值」按钮
• 工具栏：模式切换为下拉列表（逐行提取 / 高亮提取 / 歌词提取 / 朗读提取）；跟踪光标（text-cursor，光标跟随开关）、上一项/下一项（arrow-big-left/arrow-big-right）、宽度锁定（move-horizontal）、穿透锁定（lock/unlock）、隐藏背景（eye-off）、字体档位图标（heading-1~heading-5）均为语义化 lucide 图标；按钮提示默认在上方弹出、上方无空间时自动翻转到底部；穿透锁定时仅保留上一项/下一项、穿透锁定、关闭等交互按钮；设置按钮直达提词器设置页
• 透明度：设置中可调字体透明度（默认 100%）与背景透明度（默认 80%），两项均带「重置为初始值」按钮（lucide rotate-ccw）；背景色、边框与外轮廓阴影常显，透明度实时生效
• 毛玻璃：设置可开关「毛玻璃效果」并调节模糊强度（0-40px，默认 16px，带重置按钮），以 backdrop-filter 模糊并提饱和面板覆盖的底层内容；配合较低的背景透明度效果更明显，「隐藏背景」激活时自动撤除
• 主题适配：切换浅色/深色主题时窗口背景色即时更新，无需重启
• 渲染样式：内容复用主题与自定义 CSS（标题、代码块、内联格式等），仅字体大小由提词器控制；单行渲染前自动去掉行首缩进，嵌套列表等缩进行按无缩进列表/文本展示，不被误判为代码块
• 垂直居中：文本显示域内内容垂直居中，短内容不再顶置；内容超过最大高度时回退常规滚动，顶部不截断；高度上限取 min(70vh, 窗口下方可用空间)——窗口贴近屏幕下缘时长行文在窗口内滚动，而不是把窗口撑出屏幕
• 状态栏入口：右下角状态栏「打开提词器」按钮（lucide-presentation）一键打开/聚焦，可在设置中关闭
• 双击定位 + 选中：双击文本显示域，编辑器跳到捕获文本所在位置并选中对应文本（高亮模式选中匹配文本段、行模式选中整行、朗读模式滚动到正在朗读的段落并选中朗读高亮的那段文字、选中提取保留现有选择），聚焦并将选中范围滚动至视口中央
• 右键复制：右键单击文本显示域，复制捕获文本的渲染后纯文本（无 Markdown 语法），成功弹出「已复制」通知
• 状态持久化：关闭后重新打开即恢复上次窗口的位置、尺寸、模式、绑定、跟踪光标等状态；已关闭窗口的状态跨重启保留，Obsidian 重启后重开仍可恢复，已关闭的窗口不会在重启后自动弹出；窗口位置按保存值原样恢复，不再因恢复后的行文变高被抬升
• 空行回退：当前行为空时显示上一项内容（半透明占位）
• 命令：「打开提词器」「关闭所有提词器」，支持多实例

朗读（桌面端）

把当前笔记念出来：文本先过滤成「该被听见的内容」，再切成段、逐段合成、边读边取下一段，并高亮正在读的那一块。仅桌面端可用（要起子进程与本地语音服务），且只支持源码模式与实时预览模式。

开始朗读

• 三种起点：命令面板「朗读：从头读 / 从光标读 / 读选区」；正文右键菜单也有「朗读：从光标读」——右键的位置就是朗读起点
• 播放条：朗读开始后注入当前笔记顶部，提供播放/暂停、停止、段计数（n / N）与段内进度细条（分段长度不均，故不显示百分比）；服务准备与失败自动重试的进度也在这条上显示
• 命令：「朗读：播放/暂停」「朗读：停止」可绑定快捷键

文本过滤（17 项，逐项可关）

• 过滤 YAML frontmatter、围栏代码块、行内代码、表格、标题记号、引用记号、callout 标记、列表记号、#标签、链接与网址、图片嵌入、[[双链]]（保留显示名）、emoji、==高亮== 记号、星号、下划线、反斜杠转义；另有「跳过含指定短语的整行」「跳过以指定前缀开头的整行」两条整行规则
• 引用块里的代码块、表格、列表同样会被过滤；callout 的 [!note] 类型标记不会被念出来，标题与正文保留
• 过滤只影响朗读：高亮位置靠偏移映射换回原文坐标，因此过滤掉链接与 Markdown 记号后，编辑器里亮起的仍是原文对应位置

分段与首字延迟

• 按标点三级切分：句末标点（含英文句点，缩略语与小数不误切）为强边界、顿号逗号分号为弱边界（需累积够字数，默认 36 字才切）、另有段内硬上限（默认 150 字）与段落换行
• 列表项、标题、callout 标记行作为结构断点强制断开，列表不会被念成一大段；段尾标点串不会单独成段，纯标点行（如单独一行的省略号）不会送进合成
• 首段短、后续渐进加长：首段默认 15 字（首字约 3 秒出声），之后按「本段音频够不够覆盖下一段合成」逐段加长，段间不出现静默等待；设置页提供「低首字延迟（推荐）」「简单分段」两套预设，首段上限滑块可手动微调
• 「朗读：预览分段（不发声）」打印过滤后字符数、段数、预估合成与音频时长、首段文本与映射自检结果，方便对照着调参数

高亮与视图联动

• 正在读的段用编辑器原生装饰高亮（复用主题与自定义 CSS），换行、缩放、切换主题都不跑偏
• 高亮按标点细分：开启后段内高亮推进到当前分句，而不是整段一次亮起；段尾标点不纳入高亮（避免「亮完一整句再多亮一个句号」的观感）
• 自动滚动到当前段；可选光标跟随（默认关——开启后编辑器光标跟着朗读走）；光标跟随会为用户操作让路：编辑器里有选区、或你刚编辑过（约 8 秒内）时朗读不去抢光标，停手且没有选区后自动恢复——因此双击定位选中的那段文字、以及边听边改都不会被朗读打断；可选锁定编辑器

三种 TTS 引擎

• GPT-SoVITS（默认）：连接本机 api_v2.py 服务（默认 127.0.0.1:9880），可设置安装根目录（含同名双层目录自动定位）、GPT / SoVITS 权重（按同名基名成对选择，防止错配静默产出错误音色）、参考音频与参考文本、朗读语种与参考语言、输出格式（默认 ogg，体积约为 wav 的 1/8.4）与语速（0.5–2.0x）；可一键「启动本地服务」「停止本地服务」「重启本地服务」，插件会记住自己启动的进程（重启 Obsidian 后仍能认领并停止），不会去动你自己启动的服务
• Qwen3-TTS：VoiceDesign 用一段自然语言描述音色（如「温柔的年轻女声」），随每个请求发送、改完即生效；插件会生成桥接服务脚本并拉起（默认 127.0.0.1:9872），提供「一键准备环境」（装依赖、按显卡驱动装 CUDA 版 torch、写脚本）、推理设备与语种设置。注意该引擎不支持语速调节，且明显比 GPT-SoVITS 慢，适合短段与音色设计
• Windows 本地语音（零安装兜底）：直接调用 Windows 自带的 SAPI5 语音，不需要安装任何东西；可选系统语音、语速（0.5–2.0x）与音量，并带「试听」
• 切换引擎时会自动停掉由本插件启动的另外两套服务以释放显存，外部启动的服务只提示、不动

服务与诊断

• 设置页「朗读 → 服务」显示服务状态（未运行 / 启动中 / 运行中（由本插件启动）/ 运行中（外部启动，本插件不会去停它））与推理健康度（未知 / 正常 / 异常——进程活着不代表能推理，健康度由真实合成结果驱动）
• 「朗读：环境诊断」逐项检查子进程能力、Python 解释器、安装根目录、权重、参考音频、服务可达性、语音合成自检（真的合成一小段）、原生文件对话框、音频解码等，并以一条通知汇总结果
• 「朗读：查看服务日志」查看服务最近输出（中文路径按 GBK 回退解码，不会乱码）；另有「朗读：强制停止服务（按端口）」应对进程归属丢失的情况

说明与限制

• 朗读依赖本地服务与子进程，仅桌面端可用；阅读模式（Reading）不支持
• 朗读过程中编辑文档：高亮与提词器内容会随文本变更自动换算位置（不会整体偏移）；但已在合成/播放中的段仍是编辑前的文字——音频不会因编辑重新合成或重排
• 目前是段级高亮（可选推进到分句），尚无逐词高亮；也没有续读位置持久化（音频导出已支持，见下节）

生成音频文件

• 入口：选中文字 → 正文右键「朗读：生成音频文件」（无选区时菜单项不出现），或命令「朗读：生成音频文件（选区）」——与朗读共用同一条分段链与引擎，声音、语速、过滤规则与「朗读选区」完全一致
• 单文件输出：长选区按朗读规则分段逐段合成后拼成一个音频文件——多段 WAV 重打包（拷贝首段 fmt 头、PCM 顺序拼接并校验采样率/位深/声道一致），OGG 直接链接拼接，单段原样落盘；文件名 = 选区摘录 + 时间戳，同名自动加后缀不覆盖
• 保存目录：设置 → 朗读 →「生成音频文件 → 保存目录」，「浏览…」按钮打开系统资源管理器选择；留空则保存到系统下载文件夹
• 服务未运行时自动准备服务，进度显示在常驻通知上；桌面端限定

正则查找替换

当前笔记（或选区）内的正则查找、替换与替换预览。入口三条：命令面板「打开正则替换」、正文右键「正则替换…」，或编辑器里按 Ctrl+H——该热键默认属于核心命令「替换」，插件会接管它（仅在编辑器上下文、恰好 Ctrl+H 时生效；输入法组词、其他弹窗、macOS 场景一律让路；原生「替换」仍可从命令面板进入或改绑其他快捷键）。

• 查找：输入正则即时显示匹配计数，命中的文本同步点亮在编辑器里（输入即刷新、关闭即清除，不落盘、不参与撤销）；支持 i 忽略大小写 / m 多行 / s 点号匹配换行三个标志与「仅替换选区」；替换总是作用于全部匹配（匹配计数 = 替换处数，g 恒定隐含）
• 替换预览：替换前/替换后对照文本（窗口裁剪锚定第一处匹配，大文档不整篇渲染）+ 匹配列表（前 10 处）；预览里的高亮片段与列表项可点击定位到编辑器对应位置
• 替换语法：$1 / $<name> / $& / $$ 引用捕获组，\n / \t 转义；非法正则以错误行提示，不会清空已输入内容
• 写回：「一处匹配 = 一条 change」批量事务——一次事务即一步撤销，光标由 CM 自动映射，选区外内容零扰动
• 历史记录：替换成功后自动去重留存，弹窗底部可回填；上限与预览开关在设置页「正则替换」页签调整
• 打开弹窗时若已有较短选区（<100 字符）会自动转义预填为查找模式；输入框内按 Enter 直接触发替换

音乐

侧边栏歌词音乐面板（点击左侧 ribbon 音乐图标或命令「打开音乐面板」）：

歌曲播放

• 歌单面板（默认显示，三标签页）：标题栏为「歌曲列表 (数量)」标题 + 「歌单/歌词」切换按钮（lucide 图标 + 当前面板文字，歌词面板标题栏同构并显示行数计数）；顶部为共用搜索框 + 分类下拉（本地=专辑类型，在线=平台），下方三个整行等宽的分段式标签页：账号歌单（网易云登录后拉取账号全部歌单，卡片展示封面/曲目数/创建者，点开浏览曲目，歌单头部「同步到本地」批量下载整张歌单）、在线歌单（四平台实时搜索，边输边出，按关键词相关度排序，行内试听/下载；无关键词时展示四源推荐歌单首屏）、本地歌单（音频文件夹歌单，支持自定义分组与拖拽排序）
• 本地歌单自定义分组：歌曲行「分组」按钮（或右键组头）新建分组 / 归组 / 移出分组；列表按分组手风琴展示——组头为吸附式背景条（滚动时钉在列表顶部），组名跟随主题粗体与界面字体，左侧把手按住可拖动调整组序，组名 + 数量，默认全部折叠（点击展开当前组并折叠其他组，再点折叠当前组），组内行相对缩进、当前播放行左侧强调色竖条；搜索/筛选中自动展开全部有结果的组；未指派的歌进「未分组」区块；分组数据存 data.json（不产生额外文件），未建分组时保持扁平列表
• 拖拽排序：按住任意歌曲行上下拖动调整顺序（5px 阈值防误触，拖到列表上下边缘自动滚动；分组视图内只在同组内换位），顺序存 data.json 并同时决定顺序播放的切歌顺序；搜索/筛选视图下禁用拖拽
• 顺序播放按分组范围：正在播放的歌属于某个分组时，顺序播放的下一首/上一首只在该分组内循环（分组末首→分组首首，分组内仅一首时等效单曲循环）；未分组或未建分组时仍按整张歌单顺序
• 歌曲行交互：双击播放（单击留给浏览/滚动，避免误触），悬停时底色浮起 + 阴影 + 封面播放浮层淡入（不描边、指针保持默认箭头，悬停底色取不透明的主题色阶，不会透出面板底色）；封面浮层即播放/停止按钮（当前播放行悬停显示暂停图标，点即暂停/续播）；账号歌单里已下载到本地的行同样走本地播放链路
• 侧边栏歌词面板：标题栏布局与歌单面板一致；实时歌词当前行放大高亮，非当前行缩小淡化，点击任意歌词行跳转到对应时间；手动滚动即暂停自动跟随（往上翻看歌词不会被立刻拽回，换行时视口按锚点保持不动），滚回当前行完整可见后自动恢复；标题栏另有 ±2000ms 歌词偏移滑块（每首歌独立，默认 -150ms，负值=歌词提前）
• 逐字高亮（卡拉OK）：当前行按字/词变色 + 光晕；支持主流增强 LRC 的精确 <mm:ss.xx> 时间戳（如 <00:12.16>沧<00:13.00>海），无精确标记的行按行时长均分逐字
• 双语注释：原文 | 译文 竖线语法，译文以灰色小字显示在原文下方，不参与逐字高亮
• 歌单来源：设置「音频文件夹」指定 vault 内文件夹或库外 Windows 盘符绝对路径（如 D:\Music，可点「浏览」打开系统资源管理器选择），扫描其中的 MP3/FLAC/M4A/OGG 音频；音频内嵌标题/歌手/专辑/封面自动填充歌单，并探测时长与文件大小（元数据胶囊展示）
• 歌词来源：优先读取音频同目录同名 .lrc 侧车文件（在 Obsidian 中编辑保存后，正在播放的歌词即时刷新），其次音频内嵌歌词（MP3 USLT / FLAC Vorbis / M4A ©lyr / OGG comment）；外部改动音频或歌词文件后自动重读
• 在线歌词自动搜索：歌词面板检索不到歌词时自动搜索网易云 / QQ / 酷狗 / 酷我四平台，并按相关度自动应用第一份有歌词的候选（保存为侧车 .lrc 并立即显示）；全部候选都无歌词时退回候选列表手动挑选；「换歌词」按钮可随时重搜覆盖
• 播放控制（底栏两行）：第一行为「歌单/歌词」切换按钮 + 可选中复制（user-select: text）的歌名 + 「定位当前歌曲」按钮（切到本地歌单、展开所在分组并居中滚动到当前行）；第二行依次为播放模式（单曲循环 / 顺序 / 乱序，仅图标变化）、上一首 / 播放暂停 / 下一首、可拖动进度条（联动播放进度，拖动即时跳转）、时长、倍速（弹列表选择 0.5–2x）与水平音量滑条弹窗；图标按钮统一为 Obsidian 原生 clickable-icon 观感（lucide 图标、悬停背景变色、默认箭头光标）；命令面板提供播放/暂停、上一首、下一首等命令
• 后台播放与进度记忆：切换应用/最小化后继续播放；退出 Obsidian 时自动保存正在播放的歌曲与进度（播放中每 5 秒节流写盘，暂停/切歌/退出即时保存），下次启动加载该歌曲、跳到上次进度但保持暂停；歌曲已删除或不在歌单时静默跳过恢复
• 标签编辑：歌单行 hover 出现编辑按钮打开「编辑标签」（MP3/M4A 可编辑；FLAC/OGG 只读「查看标签」），支持改名与删除（vault 内进系统回收站），编辑正在播放的歌曲时标题/歌词实时刷新

歌词获取

• 在线歌词多平台：标签编辑弹窗「获取歌词」按网易云 / QQ / 酷狗 / 酷我四平台搜索当前歌曲，候选列表（含来源/时长）手动选择后导入；「获取封面」同机制提供四平台封面候选（缩略图网格）
• 试听歌曲临时歌词：无本地文件的试听歌曲同样可在线搜索并临时挂载歌词（仅本次试听有效，不写任何文件），切歌或试听结束自动失效

歌曲下载与试听

• 多平台下载：已并入侧边栏「在线歌单」标签页，实时搜索网易云 / QQ / 酷狗 / 酷我（边输入边出结果，防抖 300ms，按关键词相关度排序），下载到音频文件夹并自动内嵌歌词/封面/标题；QQ 免费歌曲无需 Cookie 可直接下载/试听（实测免登录可拿直链），VIP 歌曲需绿钻账号的有效 Cookie（新版网页登录的 Cookie 可能不被下载接口认可，以下载实测为准；「测试连接」按钮探测的是真实下载通道），网易云可选粘贴会员 Cookie 解锁 VIP 高音质（weapi/eapi 无损逐级降级）；同名文件自动追加来源后缀防覆盖
• 平台 Cookie 自动续期回写（Cookie JAR）：响应头里的 Set-Cookie 经「域名 + Cookie 名」双重白名单把关后自动合并回本地保存的平台 Cookie 并落盘——网易云 MUSIC_U / QQ qm_keyst 均为 HttpOnly，手工 F12 粘贴的静态凭证会在平台续期后作废，现在续期结果自动回收，不必再周期性重新复制；删除语义（Max-Age=0 / 已过期）不落地，平台一次清理动作不会抹掉粘贴的凭证；设置页 Cookie 行的状态徽标实时计算：「已配置（N 字符）」，发生续期后追加「已自动续期 MM-DD HH:mm」，重新粘贴或清空即失效
• 下载前试听：结果行「试听」拉取标准档音频直接播放（不写入库），经底栏主播放器播放（播放/暂停/进度/音量与歌词面板全部适用；播完按播放模式在试听队列内续播，加载失败自动跳下一首）；同歌再点命中会话内缓存秒播；设置页「释放缓存」一键清空
• 网易云客户端缓存：设置「网易云客户端缓存目录」指向客户端缓存目录后，试听与网易云歌曲下载（含账号歌单同步）优先读客户端已缓存音频，命中即离线秒开
• 推荐歌单：「在线歌单」标签页无关键词时展示四源推荐歌单（胶囊切换平台，可按平台与关键词筛选），点歌单查看歌曲并逐首下载，支持返回与刷新

来源标注：本「音乐」模块是在 LyricFlux（Copyright (c) 2026 eatgrass, fhb263）与 obsidian-lyric（Copyright (c) 2023 eatgrass）两个 MIT 项目的基础上修改而来，遵循 MIT 许可；完整声明见 LICENSE。

设置

设置界面按功能分为六个页签：

• 选择高亮：高亮选中文本出现位置开关；高亮延迟（毫秒，需 ≥200）；缩略图开关；选择检索的字符串上限滑杆（2-60，默认 30，带「恢复默认」按钮）；「高亮索引」设置项并入本页末尾（启动时默认打开高亮索引）
• 持久高亮：自定义样式的创建、编辑、删除，标签组管理与一键导入导出。表单含名称、描述、背景色（可留空）、搜索词/表达式（带正则开关）、标记开关，以及带 CSS 自动补全的代码编辑器，并提供「清空当前编辑」一键复位
• 提词器：字体（本机字体选择）、字重（跟随主题 / 100–900）、字体颜色、字体透明度（默认 100%）、背景透明度（默认 80%），均带「重置为初始值」按钮；毛玻璃效果开关与模糊强度（0-40px，默认 16px）；可折叠「文字阴影」分组（投影开关、水平/垂直偏移、模糊半径、不透明度）、「文字描边」分组（描边开关、描边方式、线宽 0-12px、颜色）与「文字渐变」分组（渐变开关、类型、范围、角度、颜色停靠点，附实时预览）；选中提取模式与状态栏按钮开关
• 音乐：音频文件夹选择器（「浏览」按钮打开系统资源管理器选择，带刷新按钮与搜索下拉，支持库外盘符绝对路径）、逐字高亮开关、底部状态栏适配开关（开启后保留 view-content 默认样式，面板底部为悬浮的应用状态栏预留安全区留白；关闭则移除该默认样式贴边铺满，适配状态栏相对布局的主题）、歌词自动滚动开关、试听缓存释放；「下载」分组为四平台可折叠 Cookie 行（启用勾选、状态徽标（含「已自动续期」时间，续期结果自动回写落盘）、「打开登录页」直达平台登录页、可点击复制的 document.cookie 代码芯片、测试连接、清除，默认收起防凭证暴露）
• 朗读：TTS 引擎下拉；朗读（预取深度、自动滚动、光标跟随、高亮按标点细分、锁定编辑器、卸载/退出时是否停止本地服务）；生成音频文件（保存目录，可「浏览…」用系统资源管理器选择，留空 = 系统下载文件夹）；分段（单段上限、弱边界阈值、分段预设、首段上限与「预览分段」）；内容过滤（17 项逐项开关 + 两条整行跳过规则）；服务与声音（GPT-SoVITS：服务状态与推理健康度、校验安装根目录、环境诊断、启动本地服务、音色选择与应用于服务、GPT / SoVITS 权重、参考音频与参考音频目录、参考音频语言、朗读语种、输出格式、语速）；服务与声音（Qwen3-TTS：服务状态、一键准备环境、校验模型目录、检查 Python 环境、推理设备、启动本地服务、音色预设与音色描述、朗读语种、环境诊断）；环境与声音（Windows 本地语音：环境状态与诊断、语音列表、语速、音量、试听）
• 正则替换：替换预览开关；历史记录上限滑杆（0-50，0 = 不记录）与「清空历史记录」按钮；「打开正则替换」入口按钮

限制

• 阅读（Reading）模式暂不支持动态高亮
• 朗读（桌面端）：依赖本地语音服务与子进程，仅桌面端可用；阅读（Reading）模式不支持；朗读期间编辑文档时，高亮与提词器会随文本变更自动换算位置，但已在合成/播放中的段仍是编辑前的文字
• 正则查找替换：作用于当前笔记或选区，暂不支持全库批量替换；预览为同步计算，超长文档上每次输入需重跑匹配，可能有轻微延迟
• 音乐：库外绝对路径音频/下载写盘为桌面端功能；移动端仅支持 vault 内音频播放，在线歌词/下载接口在移动端可能因 Referer 头被剥离而不可用

赞助

如果 Glimpse 对你有帮助，欢迎赞助支持～

致谢

感谢 @chrisgrieser（aka @pseudometa）提供的插件创意和反馈。
感谢 @chetachiezikeuzor 的插件设置界面代码，灵感来自 https://github.com/chetachiezikeuzor/highlightr-Plugin/
感谢 @eatgrass 与 @fhb263：「音乐」模块基于 LyricFlux 与 obsidian-lyric（均为 MIT 许可）修改而来
感谢 @bongho：「正则查找替换」模块的引擎与交互参考 obsidian-regex-replace 移植而来（纯逻辑引擎、替换预览与高亮渲染的原始设计均出自该项目，在此致以谢意）

---

English README

Keywords: dynamic highlighting, selection highlighting, persistent highlighting, regex queries, capture groups, custom CSS, CSS autocompletion, highlighter description, highlight index, teleprompter, cursor-linked selection, scrollbar markers, minimap, text gradient, subtitle drop shadow, scroll sync, click-through lock, document binding, group management, import/export, regex find & replace, replace preview, music player, lyrics panel, karaoke word highlighting, online lyrics, multi-platform download, note read-aloud, text-to-speech (TTS), player bar, content filtering, clause-level highlighting, audio export, generate audio file, GPT-SoVITS, Qwen3-TTS, Windows local voice

An Obsidian plugin that dynamically highlights text based on cursor selection or search query. Key features:

• Selection highlighting: instantly highlights all occurrences of the selected text (decorations appear only when at least one occurrence other than the selection is found), with scrollbar markers and a minimap
• Persistent highlighting: mark text persistently via regex/keyword queries, with capture groups, line/start/end widgets, custom CSS (a built-in code editor with CSS autocompletion), a per-highlighter description, group management, and one-click import/export
• Highlight index: auto-scans ==highlighted== text and organizes it into a sidebar index by heading hierarchy, with cursor-linked selection and keyboard navigation
• Teleprompter (desktop only): karaoke-style floating teleprompter windows that follow the document/cursor in real time, with font weights 100–900, text gradient, subtitle drop shadow, click-through lock, document binding, scroll sync, and multi-instance support
• Read aloud (desktop only): reads the current note out loud — Markdown is first filtered down to "what should be heard" by 17 rules, then split into segments at punctuation boundaries and synthesized one segment at a time while the next is fetched, with the segment being read highlighted (optionally down to the current clause); start from the top of the document, the cursor or the selection, with prefetching, pause/resume, auto-scroll, cursor following and three TTS engines (local GPT-SoVITS / Qwen3-TTS / the zero-install Windows local voice); selected text can also be exported as a single audio file in one click
• Regex find & replace: regex search, replace and replace preview inside the current note (or the selection) — summoned with Ctrl+H, a command or the context menu, with the matches lighting up live in the editor, a clickable before/after preview and match list, $1 / $<name> capture-group references and \n escapes, and per-match write-back with a single undo step (ported from obsidian-regex-replace, see the acknowledgments)
• Music: a sidebar lyrics music panel — a three-tab playlist (NetEase account playlists synced to local / online playlists searched and downloaded across four platforms / local songs with custom groups and drag ordering), karaoke word-by-word highlighting with bilingual annotations, auto-following lyrics (paused the moment you scroll manually), an audio folder (MP3/FLAC/M4A/OGG, absolute paths outside the vault supported) as the playlist, four-platform online lyrics fetching, preview-before-download and recommended playlists

Currently supports Source mode and Live Preview mode. Reading mode and the legacy editor are not supported.

Selection Highlighting

When text is selected, highlights all occurrences of the selected text:

• Case-insensitive matching across the whole document (not just the visible viewport)
• Decoration condition: decorations are applied only when at least one occurrence other than the selection itself is found; selecting text that occurs only once draws no underline and no scrollbar marker
• Current selection marked as .cm-current-string
• Other matching strings in the document marked as .cm-matched-string
• All matches include data-contents attribute with the selected string value
• Scrollbar markers / minimap: match positions shown on the scrollbar when matches are found; an optional minimap on the editor's right edge (draggable to scroll)
• Selection length cap: a "Max selection length" slider in settings (2-60, default 30) skips full-document matching for selections longer than the cap, avoiding slowdowns from huge selections

Persistent Highlighting

Define search queries with associated CSS class names and colors to create persistent highlights. Matched strings are automatically tagged with the corresponding CSS class and painted with the chosen background color (the color may be left empty — with no color picked, only the highlight class is applied and styling is left entirely to custom CSS). Each highlighter's style and color is cached in the index; match toggles can be switched on/off at any time to control how matches appear.

Each highlighter can also carry a description noting what its expression matches — display-only, never used for matching. It sits on the row below the name in settings and on the line below the expression in the list card.

Supports regex queries (enable the toggle for regex mode). In regex mode, named capture groups (e.g., (?<groupName>…)) or unnamed capture groups (e.g., (…)) can be used to highlight sub-matches with precision.

Mark Types

Each highlighter can combine multiple mark modes:

• Match: Highlight the full matched text (enabled by default)
• Line: Apply the CSS class to the entire line containing the match, enabling whole-line styling instead of word-level
• Start / End: Insert zero-width widget elements at match boundaries — use with CSS for prefix/suffix icons
• Group: In regex mode, highlight capture group sub-matches instead of the full match. Sub-matches use the highlighter's own color. Supports both named (?<name>…) and unnamed (…) capture groups. When "Group" is toggled on, full-match decoration is automatically skipped.

Custom CSS

Each highlighter can include its own CSS rules, injected into the page via CSSStyleSheet + document.adoptedStyleSheets (no <style> element is created). Renders through the editor's CodeMirror instance, supporting dark/light theme adaptation. CSS changes take effect immediately on save.

The field itself is a full CodeMirror 6 editor: syntax highlighting, auto-closing brackets and quotes, undo history, and CSS autocompletion as you type — property names, value keywords and selector tag names all suggest as you type, or on demand with Ctrl/Cmd-Space; accepting a property also inserts the colon and a space.

Group Management

Organize highlighters with groups: create, rename, delete groups; drag highlighters onto group tabs to categorize, and drag group tabs to reorder them (the default group stays first and cannot be dragged). The "Enable All / Disable All" toolbar button controls match toggles for the current group at once.

Import & Export

Supports one-click import (from clipboard JSON or a file), batch export of all highlighters with group metadata, and per-highlighter single export. The import/export dialogs use styled plain textareas: buttons sit above the box (Export to file / Export to clipboard; Import from file / Import from the paste box below), the box auto-fits its content height, and both dialogs share one visual style. Backward compatible with legacy data format (no groups → defaults to "默认").

Pandoc Fenced Divs Highlighting (:::)

{
  "Pandoc-Syntax": {
    "class": "Pandoc-Syntax",
    "color": "#77787C4A",
    "regex": true,
    "query": "::: \\{.*?\\}[\\s\\S]*?:::",
    "mark": ["match", "group"],
    "css": ""
  }
}

Highlight Index

Automatically scans the active document for ==...== wrapped Obsidian standard highlights and organizes them in a sidebar view grouped by heading hierarchy.

• Heading grouping: Highlights organized by the document's heading structure; highlights without a preceding heading grouped under the document-name group
• Level icons: h1-h6 headings display a corresponding lucide heading icon, colored by --h1-color through --h6-color CSS variables
• Ancestor breadcrumbs: Shared ancestor headings rendered as standalone rows before the first entry in each subtree, using the same icon and color styling as direct headings
• Frosted glass cards: Highlights displayed as rounded frosted glass cards with a hover-enhanced frosted effect
• Cursor-linked selection: when the editor cursor lands on a line containing a highlight, the matching card is selected and scrolled to the center of the panel
• Keyboard navigation: with the index tab focused, ↑/↓ steps to the previous/next card (counting from the currently selected one), also syncing the editor cursor and the teleprompter
• Right-click to copy: right-clicking a card copies its text and shows a Notice toast
• Anchored-document fallback: when the current page has no highlights, the index scans the document bound to the teleprompter instead; documents not open in a view are read from disk
• Auto-refresh: Scans a document on first open and re-scans on document switch (a file-open listener covers opens that don't change the active leaf, e.g. new notes or explorer clicks on the current tab); keeps previous results when the new active document has no highlights
• Command palette: Registers "打开高亮索引" command to summon the index view
• Settings toggle: "Startup auto-open highlight index" (off by default) — enables the index tab on plugin load

Teleprompter (desktop only)

Karaoke-style floating teleprompter windows that follow document content in real time — for cue cards, reading aloud, or lectures. Desktop only.

• Four content modes:
  • Line extract: static by default (extracts once when opened or when switching documents); click the "Track Cursor" button in the toolbar to enable cursor-following — the line under the cursor is auto-extracted as it moves (poll-based detection, bound or active document); wheel/buttons step up/down one line manually
  • Highlight extract: cycles through ==...== matches in order — great for scripted reading
  • Lyrics extract: shows the currently playing lyric line from the music module, advancing with playback (deduplicated by song + line index so high-frequency state pushes never re-render); prev/next/wheel seek to the adjacent lyric line's timestamp (also while paused); a dim "not playing" hint appears with no session, and "♪ intro ♪" before the first line is sung
  • Reader extract: shows the block currently being read aloud, advancing automatically with playback (down to the current clause when "clause-level highlighting" is on); the content is taken straight from the editor's original text, so Markdown markers render normally; a "not reading" placeholder appears when nothing is being read. The content is driven by the audio clock, so there is no seekable previous/next item
  • Selection extract: temporarily overrides the content with selected text, auto-restores on deselect
• Mode switch: a hand-drawn toolbar dropdown (line / highlight / lyrics / reader extract) — with a chevron-down caret, a popup menu that follows the Obsidian theme colours and highlights the active item (a native <select> opens a system-level popup that neither follows the toolbar's visibility nor the theme, so it was dropped), with the collapsed value and the popup options both centered; it closes when you click outside the teleprompter, press Esc, or move the mouse out of the window; the window's drag/wheel handlers exempt the whole dropdown (so the option list scrolls normally), and it is hidden while click-through locked
• Highlight index integration: clicking an index card binds that document and switches to highlight-extract mode; double-clicking the teleprompter text area selects the matching index card, and with scroll-sync on, prev/next steps select the corresponding card too
• Scroll sync: the button (lucide link) right of the click-through lock — when active, prev/next also syncs: line-extract mode moves the editor cursor to the previous/next line, highlight-extract mode selects the previous/next index card; hidden with the other non-interactive buttons while locked
• Document binding: pin the window to a specific document instead of following the active one; once bound, clicking the bind button again always unbinds instead of binding the currently active document
• Click-through lock: whole window becomes mouse-transparent (interactive buttons kept), never blocks the editor; the lock itself no longer changes the background — full background transparency is handled solely by the "Hide Background" button; while locked the toolbar shows only on hover and fades out on mouse-leave, and the "Track Cursor" / "Scroll Sync" buttons are hidden with the other non-interactive buttons
• Hide Background: when active, the window background becomes fully transparent (including hover/drag states) and the text automatically gets a subtitle drop shadow; when inactive, the background color stays visible with the opacity set in the settings
• Text shadow (subtitle effect): while the background is hidden, a soft drop shadow (drop-shadow, strictly beneath the text) is added to the teleprompter text; the collapsible "Text Shadow" settings group adjusts the shadow toggle, horizontal/vertical offset, blur radius and opacity (each with a reset button), defaulting to right 2px / down 2px / blur 2px / opacity 40%
• Text stroke: an outline drawn around the teleprompter text (desktop-lyrics style, easier to read over busy backgrounds, always active) — on by default, outer mode, 1px white; the collapsible "Text Stroke" settings group adjusts the stroke toggle, stroke mode, stroke width (0-12px) and stroke color (palette, "Clear" restores the default white), each with a reset button. Two modes: glyph stroke (-webkit-text-stroke + paint-order: stroke fill — the stroke sits beneath the fill and about half the width shows, the crispest option) and outer stroke (an SVG feMorphology dilation filter applied to the rendered pixels — gradient- and per-character-gradient-safe, the full width shows; the filter definition is injected into the document by the plugin and kept in sync with settings; corners can look slightly squarish at large widths). Neither mode thickens the glyphs, and both stack with the background-hidden drop shadow
• Text gradient: gradient coloring for the teleprompter text; the collapsible "Text Gradient" settings group offers a gradient toggle (overrides "Font Color"), linear/radial type, whole-text/per-character scope (per-character mode clips the gradient independently for each glyph, emoji and combined characters stay intact), gradient angle, and a color-stop editor (add/remove, position, palette, sorted by position), with a live preview sharing the teleprompter's parameters
• Width auto-fit: the background hugs its content — the visible panel uses CSS fit-content to wrap the text directly (no large blank areas even if the width state drifts due to async glitches), while the window width is computed live from the content's widest line and font size, clamped to the viewport so long text wraps; the toolbar doesn't participate in width calculation and wraps when the window is narrower; fitting is triggered redundantly (render completion + a ResizeObserver for late changes like font loads or theme switches + viewport resize + a once-per-second reconciliation) so any drift self-corrects; the window root never intercepts mouse events (a drifted transparent area can't block the editor), and the resize handle always sits on the visible panel's right edge; while width is locked (or being dragged) the panel fills the chosen width — drag the right edge to resize and it auto-locks, staying draggable while locked with the new width inherited as the locked value
• Drag snapping: snaps to viewport edges and center lines with guide overlays
• Stable position: the window's position is yours — content-height changes (line changes, font-size changes, width re-fitting) never move it; after a restart it comes back exactly where it was saved instead of being lifted by taller restored content
• Font size: cycles 32/40/50/64/80px
• Font / weight / color: adjustable in settings — the body font uses a local-font picker modal (Cancel/OK pinned to the header row so no scrolling is ever needed; queryLocalFonts() enumeration, falling back to canvas measurement over a candidate list; multi-select with drag-to-reorder priority — the first available font wins, missing ones fall through; search, self-preview, and custom-font input included), font weight is a dropdown (follow theme / 100–900), and text color is a palette applied live ("Clear" restores the theme default); each has a "Reset to initial value" button
• Toolbar: the mode toggle is a dropdown (line / highlight / lyrics / reader extract); track cursor (text-cursor, cursor-following toggle), prev/next (arrow-big-left/arrow-big-right), width lock (move-horizontal), click-through lock (lock/unlock), hide background (eye-off), and font-size slot icons (heading-1~heading-5) all use semantic lucide icons; button tooltips default to popping above and flip below only when there is no room above; when locked, only interactive buttons (prev/next, lock, close) remain; the settings button jumps straight to the teleprompter settings page
• Opacity: font opacity (default 100%) and background opacity (default 80%) adjustable in settings, each with a "Reset to initial value" button (lucide rotate-ccw); the background color, border and outline shadow are always visible and the opacity applies live
• Frosted glass: a "Frosted glass" toggle and a blur-strength slider (0-40px, default 16px, with a reset button) in settings; uses backdrop-filter to blur and saturate the content beneath the panel — best paired with a lower background opacity, and it is removed automatically while "Hide Background" is active
• Theme adaptation: the window background updates instantly when toggling light/dark themes — no restart needed
• Theme / custom CSS styling: content reuses theme and user CSS snippets (headings, code blocks, inline formatting); only font size is controlled by the teleprompter. Leading tabs/spaces are stripped before rendering a single line, so indented content (e.g. nested list items) renders as unindented list/text instead of a code block
• Vertical centering: content is vertically centered within the text display area — short content no longer sits at the top; when content exceeds the max height it falls back to normal scrolling (top never clipped); the height cap is min(70vh, space below the window), so long content near the bottom of the screen scrolls inside the window instead of pushing it off-screen
• Status bar entry: "Open teleprompter" button in the bottom-right status bar (lucide-presentation) for one-click open/focus, toggleable in settings
• Double-click to jump & select: double-click the text area to take the editor to the captured text's position and select the matching text (the match segment in highlight-extract mode, the whole line in line mode, in reader-extract mode it scrolls to the segment being read and selects exactly the text the reader highlights, the existing selection preserved in selection-extract mode), focus the editor, and scroll the selection to the center of the viewport
• Right-click to copy: right-click the text area to copy the rendered plain text of the captured content (no Markdown syntax); a "已复制" notice confirms success
• State persistence: closing a window and reopening it restores its position, size, mode, binding, track-cursor toggle, and more; closed-window states survive an Obsidian restart (reopening still restores them), and closed windows never auto-reopen on restart; the window position comes back exactly as saved instead of being lifted by taller restored content
• Empty-line fallback: shows previous item's text (half-opacity placeholder) when the current line is empty
• Commands: "打开提词器" and "关闭所有提词器", multiple instances supported

Reading aloud (desktop only)

Reads the current note out loud: the text is first filtered down to "what should be heard", then split into segments, synthesized one segment at a time while the next is fetched, with the segment being read highlighted. Desktop only (it spawns child processes and talks to local speech services), and it supports only Source mode and Live Preview mode.

Starting a session

• Three starting points: the command palette's "Read aloud: from top / from cursor / read selection"; the editor's right-click menu also offers "Read aloud: from cursor" — the clicked position is the starting point
• Player bar: injected at the top of the current note once reading starts, with play/pause, stop, a segment counter (n / N) and a thin in-segment progress bar (segment lengths are uneven, so no percentage is shown); service preparation and automatic retry-after-failure progress also appear on this bar
• Commands: "Read aloud: play/pause" and "Read aloud: stop" can be bound to hotkeys

Content filtering (17 toggles)

• Filters YAML frontmatter, fenced code blocks, inline code, tables, heading markers, quote markers, callout markers, list markers, #tags, links and URLs, image embeds, [[wikilinks]] (keeping the display name), emoji, ==highlight== markers, asterisks, underscores and backslash escapes; two whole-line rules — "skip lines containing a given phrase" and "skip lines starting with a given prefix" — are also available
• Code blocks, tables and lists inside blockquotes are filtered as well; a callout's [!note] type marker is never read out, while its title and body are kept
• Filtering affects reading only: highlight positions are mapped back to original-document coordinates by an offset map, so after links and Markdown markers are filtered out the editor still lights up the corresponding original positions

Segmentation & first-word latency

• Three-level splitting at punctuation: end-of-sentence punctuation (including the English period; abbreviations and decimals are never split) is a strong boundary, enumeration commas, commas and semicolons are weak boundaries (they need enough accumulated characters — the default is 36 characters before splitting), plus an in-segment hard cap (default 150 characters) and paragraph breaks
• List items, headings and callout marker lines act as structural breaks and force a split, so a list is never read as one giant segment; a trailing punctuation run never becomes its own segment, and pure punctuation lines (e.g. an ellipsis alone on a line) are never sent to synthesis
• Short first segment, then progressively longer: the first segment defaults to 15 characters (first word out in about 3 seconds), after which segments grow one by one based on whether "this segment's audio is long enough to cover synthesizing the next one", so there is never a silent wait between segments; the settings page offers "Low first-word latency (recommended)" and "Simple segmentation" presets, and the first-segment cap slider can be fine-tuned by hand
• "Read aloud: preview segments (silent)" prints the filtered character count, segment count, estimated synthesis and audio durations, the first segment's text and a mapping self-check, making it easy to tune parameters against

Highlighting & view linkage

• The segment being read is highlighted with the editor's native decorations (reusing the theme and custom CSS), staying accurate across line changes, zooming and theme switches
• Clause-level highlighting: when on, the segment's highlight advances to the current clause instead of lighting up the whole segment at once; trailing punctuation is not included in the highlight (avoiding the "the whole sentence lights up and then one more period lights up" feel)
• Auto-scroll to the current segment; an optional cursor follow (off by default — when on, the editor cursor follows the reading); cursor follow defers to whatever you are doing: it never takes the cursor while the editor has a selection, or for about 8 seconds after your last edit, then resumes automatically once you stop and nothing is selected — so both the text selected by a double-click and any editing-while-listening stay uninterrupted; an optional lock editor

Three TTS engines

• GPT-SoVITS (default): connects to a local api_v2.py service (default 127.0.0.1:9880), with an install root setting (including automatic detection of a same-named double-nested directory), GPT / SoVITS weights (selected in pairs by matching base name, preventing mismatches from silently producing the wrong voice), reference audio and reference text, reading language and reference language, output format (default ogg, roughly 1/8.4 the size of wav) and speed (0.5–2.0x); one-click "Start local service", "Stop local service" and "Restart local service" are provided, and the plugin remembers the process it started (it can still claim and stop it after an Obsidian restart) — it never touches a service you started yourself
• Qwen3-TTS: VoiceDesign takes a natural-language description of the voice (e.g. "a gentle young female voice"), sent with every request and taking effect as soon as it is edited; the plugin generates a bridge service script and launches it (default 127.0.0.1:9872), offering one-click environment setup (installing dependencies, installing the CUDA build of torch for your GPU driver, writing the script), plus inference device and language settings. Note that this engine does not support speed adjustment and is noticeably slower than GPT-SoVITS, suiting short segments and voice design
• Windows local voice (zero-install fallback): calls Windows' built-in SAPI5 voices directly with nothing to install; you can pick the system voice, speed (0.5–2.0x) and volume, and it has a "Preview" button
• Switching engines automatically stops the other two services that this plugin started to free VRAM; externally started services are only reported, never stopped

Service & diagnostics

• The settings page's "Read aloud → Service" shows service status (not running / starting / running (started by this plugin) / running (started externally — this plugin will not stop it)) and inference health (unknown / OK / failing — a live process does not mean it can infer; health is driven by real synthesis results)
• "Read aloud: environment diagnostics" checks child-process capability, the Python interpreter, the install root, weights, reference audio, service reachability, a speech synthesis self-test (it really synthesizes a short clip), native file dialogs, audio decoding and more, then summarizes the results in a single notice
• "Read aloud: view service log" shows the service's latest output (Chinese paths fall back to GBK decoding instead of garbling); "Read aloud: force-stop service (by port)" covers the case where process ownership has been lost

Notes & limitations

• Reading aloud depends on local services and child processes and is desktop only; Reading mode is not supported
• Editing the document while reading: the highlight and the teleprompter text re-map to the current document automatically (they no longer drift as a whole); segments already synthesised or playing still read the pre-edit text, though — the audio is never re-synthesised or re-ordered because of an edit
• Highlighting is currently segment-level (optionally advancing to the clause), with no word-level highlighting yet; there is also no resume-position persistence (audio export is now supported — see below)

Generate audio file

• Entry points: select text → right-click "朗读：生成音频文件" (the item only appears with a non-empty selection), or the command "朗读：生成音频文件（选区）" — it shares the reader's exact segmentation chain and engine, so the voice, speed and filtering match "read the selection" one to one
• Single-file output: long selections are segmented and synthesized per the reading rules, then joined into one audio file — multiple WAV segments are repacked (the first segment's fmt header is copied verbatim, PCM streams concatenated, with sample rate / bit depth / channel count verified), OGG is concatenated by chaining, and a single segment is written untouched; the file name is the selection excerpt + a timestamp, and duplicates get a numeric suffix instead of overwriting
• Save folder: Settings → Reading aloud → "Generate audio file → Save folder", with a "Browse…" button that opens the system file explorer; when left empty the file goes to the system Downloads folder
• The service is prepared automatically when it is not running, and progress shows on a persistent notice; desktop only

Regex Find & Replace

Regex search, replace and replace preview inside the current note (or the selection). Three entry points: the command palette's "打开正则替换", the context menu's "正则替换…", or Ctrl+H in the editor — that hotkey belongs to the core "Replace" command by default, and the plugin takes it over (only in the editor context and for exactly Ctrl+H; IME composition, other modals and macOS all fall through; the native "Replace" remains reachable from the command palette or by rebinding).

• Search: the match count updates as you type, and the hits light up live inside the editor (refreshed on every keystroke, cleared when the dialog closes — nothing persisted, nothing in the undo stack); i / m / s flags and a "selection only" toggle are supported; replacement always applies to every match (match count = replacements made, with g permanently implied)
• Replace preview: a before/after comparison (window-clipped and anchored on the first match, so large documents are never rendered whole) plus a match list (first 10 entries); both the highlighted fragments and the list items are clickable to jump to that position in the editor
• Replacement syntax: $1 / $<name> / $& / $$ capture-group references and \n / \t escapes; an invalid regex is reported on an error line without clearing what you typed
• Write-back: replacements are dispatched as "one match = one change" in a single transaction — one undo step, the cursor mapped by CM, and content outside the selection untouched
• History: successful replacements are stored deduplicated and can be recalled at the bottom of the dialog; the cap and the preview toggle live in the settings' "正则替换" tab
• Opening the dialog with a short selection (<100 characters) pre-fills the pattern with the escaped selection; Enter in either input triggers the replace

Music

A sidebar lyrics music panel (click the music ribbon icon or run the "Open music panel" command):

Playback

• Playlist pane (default, three tabs): the header holds a "Songs (count)" title + the "playlist/lyrics" switch button (lucide icon + the current pane's name; the lyrics pane's header mirrors it and shows a line count); below is the shared search box + category dropdown (local = album type; online = platform), then three full-width equal-width segmented tabs: Account playlists (after a NetEase login it lists every playlist on the account — cover, track count and creator on each card; open one to browse its tracks, and its header has "sync to local" to batch-download the whole playlist), Online playlists (real-time four-platform search streaming as you type, sorted by keyword relevance, with inline preview/download; with no keyword it shows the four-source recommended-playlist home screen), Local songs (the audio-folder playlist, with custom groups and drag ordering)
• Custom groups for the local playlist: a "group" button on each song row (or right-click a group header) creates a group, assigns the song, or ungroups it; the list renders grouped accordion sections — sticky header bars pinned to the top of the list while scrolling, group names using the theme's bold styling and the interface font, a handle on the left that reorders groups by dragging, name + count, all collapsed by default (clicking expands that group and collapses the others, clicking again collapses it), rows indented inside the group with an accent bar on the playing row; while searching/filtering all matching groups render expanded; unassigned songs fall into an "ungrouped" section; group data lives in data.json (no extra files), and with no groups defined the list stays flat
• Drag to reorder: press and drag any song row up or down to reorder it (a 5px threshold avoids accidental drags; dragging near the list's edges auto-scrolls; inside a group view the move stays within the same group); the order is stored in data.json and also drives the sequential play mode's track order; dragging is disabled while searching/filtering
• Sequential playback scoped to the group: when the playing song belongs to a custom group, sequential mode's next/previous only cycles within that group (the group's last song wraps to its first; a one-song group behaves like single loop); songs outside any group, or with no groups defined, keep the whole-playlist order
• Song row interaction: double-click to play (single clicks are left for browsing/scrolling), with a hover state of a raised background + shadow + cover play overlay fading in (no outline, default arrow cursor, and the hover background uses an opaque step of the theme scale so the panel never shows through); the cover overlay is the play/stop button (the playing row shows a pause icon on hover — click to pause/resume), and rows in account playlists that already exist locally use the same local playback path
• Sidebar lyrics panel: header layout mirrors the playlist pane; the current line is enlarged and highlighted while the others shrink and fade, and clicking any line seeks to it; scrolling manually pauses auto-follow (scrolling up to read earlier lyrics is never yanked back, and line changes keep the viewport pinned by anchor), resuming automatically once the current line is fully back in view; the header also carries a ±2000ms lyric offset slider (per song, default -150ms, negative = lyrics earlier)
• Karaoke word highlighting: the current line is colored word by word with a glow; supports precise <mm:ss.xx> timestamps from enhanced LRC (e.g. <00:12.16>Words<00:13.00>more), with even distribution as fallback
• Bilingual annotations: original | translation pipe syntax renders the translation as small gray text below the original, excluded from word highlighting
• Playlist source: point the "Audio folder" setting at a vault folder or an absolute Windows path outside the vault (e.g. D:\Music; a "Browse" button opens the system file explorer), and MP3/FLAC/M4A/OGG files are scanned; embedded title/artist/album/cover populate the playlist, and duration and file size are probed and shown as metadata pills
• Lyrics sources: a same-named .lrc sidecar file next to the audio takes priority (editing it in Obsidian refreshes the playing lyrics immediately), falling back to embedded lyrics (MP3 USLT / FLAC Vorbis / M4A ©lyr / OGG comment); external changes to the audio or lyrics file are re-read automatically
• Automatic online lyric search: when the lyrics pane finds no lyrics it searches the four platforms (NetEase / QQ / Kugou / Kuwo) and applies the first candidate that has lyrics in relevance order (saved as a sidecar .lrc and shown immediately); if none has lyrics it falls back to the manual candidate list, and the "replace lyrics" button re-searches at any time
• Playback controls (two-row bottom bar): row 1 holds the "playlist/lyrics" switch + the selectable/copyable song title + a "locate current song" button (switches to the local playlist, expands the group and center-scrolls to the row); row 2 holds the play mode (single loop / sequential / shuffle, icon-only changes), prev / play-pause / next, a draggable seek bar (synced with playback, instant scrubbing), the time label, speed (0.5–2x from a popup list) and a horizontal volume slider popup; icon buttons follow Obsidian's native clickable-icon look (lucide icons, hover background tint, default arrow cursor); the command palette offers play/pause, previous and next
• Background playback & resume: music keeps playing when the app is minimized or in the background; on quit the playing track and position are saved automatically (throttled to every 5s while playing, immediate on pause/track change/quit) and the next launch loads that track, seeks to the saved position and stays paused; restore is silently skipped if the file is gone
• Tag editing: hover a playlist row to edit tags (MP3/M4A writable; FLAC/OGG read-only viewer), rename and delete (vault files go to the system trash); the playing song's title/lyrics refresh live

Lyrics fetching

• Multi-platform online lyrics: the tag editor's "Fetch lyrics" button searches NetEase / QQ / Kugou / Kuwo for the current song and presents a candidate list (source + duration) to import manually; "Fetch cover" works the same way with a thumbnail grid
• Temporary lyrics for previews: previewed songs (no local file) can also search online and mount lyrics in memory for that preview session only (nothing written to disk), clearing on track change or preview end

Download & preview

• Multi-platform download: merged into the sidebar's "Online playlists" tab, searching NetEase / QQ / Kugou / Kuwo in real time (results stream in as you type, debounced 300ms, sorted by keyword relevance) and saving to the audio folder with lyrics/cover/title embedded automatically; free QQ songs download/preview without any cookie (verified working), VIP songs require a green-diamond account's valid cookie (newer web-login cookies may be rejected by the download API — the "Test connection" button probes the real download channel instead); an optional NetEase VIP cookie unlocks lossless quality (weapi/eapi with fallbacks); duplicate filenames get a source suffix
• Automatic platform-cookie renewal write-back (Cookie JAR): Set-Cookie response headers pass a double whitelist (hosts + cookie names) and are merged back into the locally saved platform cookies automatically — NetEase's MUSIC_U and QQ's qm_keyst are both HttpOnly, so hand-pasted static credentials get invalidated by server-side renewals; renewals now flow back on their own instead of forcing periodic re-copying; deletion semantics (Max-Age=0 / expired) never land, so one platform-side cleanup sweep can't wipe the pasted credential; each cookie row's status badge in the settings computes live — "configured (N chars)", gaining "auto-renewed MM-DD HH:mm" once a renewal lands, invalidated by re-pasting or clearing
• Preview before download: the "preview" button on each result pulls the standard-quality audio and plays it immediately (nothing written to the vault) through the bottom-bar main player (play/pause, seek, volume and the lyrics pane all apply; when it ends it continues within the preview queue per the play mode, skipping failures automatically); a second click on the same song replays instantly from the in-session cache, and the settings page can flush the cache
• NetEase client cache: point the "NetEase client cache folder" setting at the desktop client's cache and previews plus NetEase downloads (including account-playlist sync) read the cached audio first — instant and offline on a hit
• Recommended playlists: with no keyword the "Online playlists" tab shows four-source recommendations (capsule switcher, filterable by platform and keyword); open a playlist to download songs one by one, with back navigation and refresh

Attribution: this Music module is modified from two MIT projects — LyricFlux (Copyright (c) 2026 eatgrass, fhb263) and obsidian-lyric (Copyright (c) 2023 eatgrass) — and is distributed under the MIT license; see LICENSE for the full notices.

Settings

The settings dialog is organized into six tabs:

• Selection: toggle highlighting all occurrences of the selected text; highlight delay in milliseconds (≥200); minimap toggle; a "Max selection length" slider (2-60, default 30, with a "Restore default" button); the "Highlight index" setting now lives at the end of this tab ("Startup auto-open highlight index")
• Persistent: create, edit, and delete highlighters, group management, one-click import/export. The form covers name, description, background color (may be left empty), search term/expression (with a regex toggle), mark toggles and a code editor with CSS autocompletion, plus a one-click "Clear current edit" reset
• Teleprompter: font (local-font picker), font weight (follow theme / 100–900), font color, font opacity (default 100%) and background opacity (default 80%), each with a "Reset to initial value" button; a frosted-glass toggle and blur strength (0-40px, default 16px); collapsible "Text Shadow" group (shadow toggle, horizontal/vertical offset, blur radius, opacity), "Text Stroke" group (stroke toggle, stroke mode, width 0-12px, color) and "Text Gradient" group (gradient toggle, type, scope, angle, color stops, with a live preview); selection-extract mode and status-bar button toggles
• Music: audio folder picker (a "Browse" button opens the system file explorer to pick a folder, plus a refresh button and search dropdown; absolute Windows paths supported), karaoke highlighting toggle, a bottom-status-bar adapt toggle (when on, the Obsidian default view-content styles are kept so the panel bottom reserves safe-area space for the floating app status bar; when off, those defaults are removed and the panel fills edge-to-edge — for themes with a relative status bar), lyrics auto-scroll toggle, preview cache flush; the "Download" group holds collapsible per-platform cookie rows (enable checkbox, status badge — including the "auto-renewed" time, with renewals written back and persisted automatically — an "Open login page" button that jumps straight to the platform's login page, a click-to-copy document.cookie code chip, test connection, clear — collapsed by default); platform-priority drag ordering was removed (results are relevance-sorted) and drag-to-reorder platform priority
• Reading aloud: the TTS engine dropdown; Read aloud (prefetch depth, auto-scroll, cursor follow, clause-level highlighting, lock editor, whether to stop local services on unload/quit); Generate audio file (the save folder, pickable via the system file explorer with "Browse…", empty = the Downloads folder); Segmentation (max segment length, weak-boundary threshold, segmentation presets, first-segment cap and "Preview segments"); Content filtering (17 individual toggles + two whole-line skip rules); Service & voices (GPT-SoVITS: service status and inference health, verify install root, environment diagnostics, start local service, voice selection and apply to service, GPT / SoVITS weights, reference audio and reference audio folder, reference audio language, reading language, output format, speed); Service & voices (Qwen3-TTS: service status, one-click environment setup, verify model folder, check Python environment, inference device, start local service, voice presets and voice description, reading language, environment diagnostics); Environment & voices (Windows local voice: environment status and diagnostics, voice list, speed, volume, preview)
• Regex replace: the replace-preview toggle; a history-limit slider (0-50, 0 = off) and a "Clear history" button; an "Open regex replace" entry button

Limitations

• Reading mode does not support dynamic selection highlighting
• Reading aloud (desktop only): depends on local speech services and child processes, so it is desktop-only; Reading mode is not supported; while reading, editing the document does not re-order segments already queued for synthesis, so highlight positions can drift
• Regex find & replace: operates on the current note or the selection only — vault-wide batch replacement is not supported; the preview computes synchronously, so on very long documents every keystroke re-runs the match and a slight delay is possible
• Music: audio outside the vault (absolute paths) and download-to-disk are desktop-only; on mobile only in-vault audio playback works, and online lyrics/download APIs may be unavailable because the Referer header is stripped

Sponsorship

If Glimpse helps you, consider supporting the author:

PayPal

Acknowledgments

Thanks to @chrisgrieser (aka @pseudometa) for the plugin concept and feedback.
Thanks to @chetachiezikeuzor for the settings UI code, inspired by https://github.com/chetachiezikeuzor/highlightr-Plugin/
Thanks to @eatgrass and @fhb263: the Music module is modified from LyricFlux and obsidian-lyric (both MIT licensed)
Thanks to @bongho: the Regex find & replace module is ported with reference to obsidian-regex-replace — the original designs of its pure-logic engine, replace preview and highlight rendering all come from that project, with gratitude