给 Claude、Codex、ChatGPT 下指令,写代码、写文档、回消息:说出来比打字快三倍,而且声音只留在你的 Mac 上。 Prompt your coding agent, draft a doc, answer a message: speaking is three times faster than typing, and your voice never leaves your Mac.
说话时,屏幕底部有一条跟着声音跳动的波形。While you talk, a live waveform sits at the bottom of the screen.
松开键,文字就粘贴到你刚才所在的输入框里。Let go, and the text is pasted where you were typing.
Cadabra 住在菜单栏里。在任何 App 里按住一个键说话,松开,文字就出现在光标处,中英混着说也行。Cadabra lives in your menu bar. In any app, hold a key, talk, let go, and the text appears at your cursor, Chinese and English mixed if you like.
几秒钟的话,识别只要 约 50 毫秒。松开键,字几乎同时出现。A few seconds of speech is recognized in about 50 ms. The text is there almost as you let go.
“这个 PR 先别 merge” 照样听得懂。也支持粤语、日语和韩语,会自己判断你在说哪种。“这个 PR 先别 merge” works as said. Cantonese, Japanese and Korean too, detected automatically.
在原处改掉听错的词,它就记住。只学真的听错了的(读音相近),不会把你改写的内容当成错误。Fix a misheard word where it landed and Cadabra remembers. It only learns real mishearings (they sound alike), never your rewording.
打开后,用本机的 Qwen 模型按上下文改正同音字和专有名词,参考你的常用词。同样不联网。Turn it on and a local Qwen model fixes sound-alikes and names from context, using your vocabulary. Still offline.
只粘贴到你说话时所在的输入框。中途切了窗口、或光标在密码框里,就只复制到剪贴板。你原来的剪贴板会放回去。Text goes only into the field you were in. Switched windows, or a password field? It's copied instead. Your own clipboard is put back.
除了默认的样式,还有星夜、纸墨、糖果,以及 Dracula、Nord、Solarized 这些写代码时熟悉的配色。Besides the default: Starlight, Paper and Candy, plus the editor palettes you know, like Dracula, Nord and Solarized.
识别模型 SenseVoice-Small 就打包在 App 里。没有账号,没有云端,不上传任何东西。The SenseVoice-Small speech model ships inside the app. No account, no cloud, nothing uploaded.
# 自己看看 Cadabra 有没有网络连接# See for yourself: Cadabra's network connections lsof -i -a -p $(pgrep -x Cadabra) # 什么都不输出,就是没有连接# No output means no connections
设置 › 通用 › 主题,选了马上生效,菜单、录音提示条和设置窗口一起换。Settings › General › Theme changes the menu, the recording bar and the windows at once.
需要 macOS 15 或更新、Apple 芯片的 Mac。Needs macOS 15 or later on an Apple silicon Mac.
从 Releases 下载 Cadabra-x.y.z.dmg,把 Cadabra 拖进「应用程序」。Get Cadabra-x.y.z.dmg from Releases and drag Cadabra into Applications.
如果 macOS 提示“无法验证”,到 系统设置 › 隐私与安全性,点 Cadabra 旁边的「仍要打开」。只需要一次。If macOS says it can't verify the app, go to System Settings › Privacy & Security and click Open Anyway next to Cadabra. Only once.
麦克风(听你说话)和辅助功能(监听快捷键、把字粘贴到光标处)。欢迎窗口会一步步带你完成。Microphone (to hear you) and Accessibility (for the shortcut and pasting at your cursor). The welcome window walks you through it.
松开就输入。或者轻按一下开始,说完再按一下结束。快捷键可以换。Let go to type it. Or tap once to start and again to finish. The key can be changed.
不要。Cadabra 免费开源。觉得好用的话,欢迎在 GitHub 上赞助或请我喝杯咖啡。No. Cadabra is free and open source. If it helps you, sponsoring on GitHub or a coffee is always appreciated.
不能,需要 Apple 芯片(M1 或更新)。No, it needs Apple silicon (M1 or newer).
能。默认会用电脑或摄像头的麦克风录音,耳机继续放音乐。用耳机的麦克风会让它切到通话模式,开始慢好几秒、声音也会变闷;要用的话可以在设置里指定。Yes. By default it records with the Mac's or camera's microphone and leaves your AirPods playing; using their microphone switches them to call mode, which delays the start and muffles the sound. You can still pick it in Settings.
不用。不下载也能正常听写。纠错模型比较大(4B 版约 2.7 GB,35B 版约 22 GB),适合常说专有名词的人;不用时它会自己从内存里卸载。No; dictation works without it. The model is large (about 2.7 GB for 4B, 22 GB for 35B) and helps most if you say a lot of names and jargon. It unloads itself from memory when idle.
对,1.2 起改名叫 Cadabra。装上新版后,设置、历史记录、学到的词和模型都会自动搬过来。Yes, until 1.2. Install Cadabra and your settings, history, learned words and models move over by themselves.