系统声音，直接采集。System audio, captured directly.
macOS 用 Core Audio Process Tap，Windows 用 WASAPI loopback。不需要 BlackHole，也不需要改音频路由。Core Audio Process Tap on macOS, WASAPI loopback on Windows. No BlackHole, no routing setup.
CAPTURE · LOOPBACK离线优先 · 开源 · MITOffline-first · Open source · MIT
所闻，即所见。Hear it. Read it.
监听系统声音输出，把外语播客、视频与会议实时变成「原文 + 中文译文」双语悬浮字幕。识别与翻译都在本机完成。 AudioScript listens to your system audio and turns foreign podcasts, videos, and meetings into live bilingual subtitles — source text plus Chinese translation, recognized and translated entirely on your machine.
主窗口 + 悬浮字幕窗 · 界面由 HTML/CSS 重建Main window + floating overlay · rebuilt in HTML/CSS
100%本地识别与翻译Local recognition & translation
0音频上传Audio uploaded
≈58MB内置模型Bundled model
70–77ms单句翻译Per-sentence translation
翻译延迟为 release 构建、M 系列 CPU、beam=1 实测。Translation latency measured on release builds, M-series CPU, beam=1.
为长时间「听」而设计。Built for long listening.
该有的都有，不该打扰的不打扰。Everything it needs. Nothing in the way.
流式双语字幕。Streaming bilingual subtitles.
partial 实时刷新，句末定稿；原文在上，译文在下。英语 ↔ 中文，逐句本地翻译。Partials update live and freeze at sentence end. Source above, translation below. English ↔ Chinese, translated sentence by sentence.
STREAM · EN ↔ ZH悬浮窗，随时让路。An overlay that gets out of the way.
始终置顶；锁定后鼠标完全穿透，不抢焦点。字号、透明度、行数与发光边框都能调。Always on top; locked mode clicks through and never steals focus. Tune size, opacity, lines, and glow.
OVERLAY · CLICK-THROUGH托盘常驻，干净导出。Tray-resident, clean export.
关窗不退出、登录时启动；定稿段落带时间戳导出为 Markdown，原文与译文成对保留。Close the window and keep listening. Export final segments to Markdown with timestamps, source and translation kept together.
TRAY · MARKDOWN字幕，应该长在屏幕上。Subtitles that belong on screen.
调一调，看看它在视频上的样子。Tune it. See how it sits on video.
OVERLAY始终深色Always dark
发光边框Glow border
只有背景透明，文字始终不透明。锁定态下鼠标完全穿透。Only the background is transparent; text never is. Locked mode clicks straight through.
模型在本地，声音不出门。Local models. Private audio.
识别与翻译都在本机完成。没有账号，没有音频上传链路。Recognition and translation run locally. No account, no audio upload path.
运行时无需网络。可选模型下载与检查更新需要联网，除此之外音频与文本不会离开本机。The runtime needs no network. Optional model downloads and update checks do; other than that, audio and text never leave your machine.
内置语音识别Bundled ASR
58 MB中英双语流式识别，随应用分发，开箱即用。Bilingual streaming recognition, shipped with the app.
Apache-2.0高精度语音识别High-accuracy ASR
≈190 MB更大的双语模型，识别更准，占用更多内存与 CPU。下载包 511 MB。A larger bilingual model. More accurate, heavier on memory and CPU. 511 MB download.
Apache-2.0 · 按需下载optional内置机器翻译Bundled translation
≈164 MBen → zh 与 zh → en 双向轻量模型，延迟低，随应用分发。Lightweight en → zh and zh → en models with low latency, shipped with the app.
Apache-2.0 / CC-BY-4.0把它放到你的桌面上。Put it on your desktop.
安装包内含内置模型，首次启动无需联网。The installer includes the bundled models. First launch needs no network.
v0.2.5·2026-10-19·全部版本All releases
macOS当前系统Your system
当前不提供 Intel Mac 安装包。No Intel Mac build at the moment.
Windows当前系统Your system
提供 NSIS 安装包与 MSI。NSIS installer and MSI available.
Linux当前系统Your system
二等支持；GNOME Wayland 下无法保证窗口置顶。Second-tier support; top-most is unavailable on GNOME Wayland.
国内网盘下载Mainland China mirror / BAIDU NETDISK
百度网盘提供全部安装包，提取码 16bk。 All installers are available on Baidu Netdisk. Extraction code 16bk.
首次开始转写需要加载模型，约 1.25–2.2 秒后出字。The first transcription loads the models, with text appearing after about 1.25–2.2 seconds.
还有一些问题。A few questions.
音频会上传吗？Is my audio uploaded?
不会。识别与翻译都在本机完成，音频与文本不上传。可选模型下载与检查更新需要联网。No. Recognition and translation run locally; audio and text are not uploaded. Optional model downloads and update checks need network.
支持哪些语言？Which languages are supported?
英语 ↔ 中文，需要手动选择翻译方向；不支持自动语言检测。English ↔ Chinese. You choose the direction; there is no automatic language detection.
需要安装 BlackHole 之类的虚拟声卡吗？Do I need BlackHole or another virtual audio driver?
不需要。macOS 使用 Core Audio Process Tap，Windows 使用 WASAPI loopback。首次采集系统声音时需授予屏幕与系统音频录制权限。No. macOS uses Core Audio Process Tap and Windows uses WASAPI loopback. The first system-audio capture requires Screen & System Audio Recording permission.
为什么开始后要等一两秒才出字？Why does the first line take a second or two?
首次开始需要加载识别与翻译模型，实测约 1.25–2.2 秒；之后进入流式输出。The first run loads the ASR and translation models, measured at about 1.25–2.2 seconds. After that it streams.