Friday

开源 · 免费 · 只在你的 Mac 上跑Open source · Free · Runs only on your Mac

一句话里中英来回切，
也不丢词、不乱翻译。Dictation for people who switch between Chinese and English mid‑sentence.

按住一个键说话，松手，文字直接出现在光标处。不上云，不要账号，没有统计。Hold a key, speak, release. The text lands at your cursor. No cloud, no account, no telemetry.

项目测试集里的一句：A sentence from the project's test set:
我让 Codex review 一下这个 pull request，然后部署到 production

英文还是英文，中文还是中文，不丢词，不翻译。English stays English, Chinese stays Chinese. Nothing dropped, nothing translated.

下载最新版Download latest release 在 GitHub 查看源码View source on GitHub

需要 Apple 芯片 Mac、macOS 14+。目前是未签名预览版，第一次打开要多点一步，看安装。Requires an Apple Silicon Mac on macOS 14+. Builds are unsigned previews, so first launch needs one extra step. See Install.

和别的语音输入有什么不一样What's different

本地中英混说的工具不止 Friday 一个（比如 TypeAloud 也做）。下面是 Friday 自己的做法，哪些别人也有，我们直接写出来。Friday isn't the only on-device tool for mixed Chinese and English (TypeAloud does it too). Here's how Friday approaches it, and where others offer the same thing.

语言总是指定，混说结果公开Language is always set, and the mixed-speech results are public

Friday 不让模型自己猜语言：让 Qwen3-ASR 自动判断时，它会把混说里的中文翻成英文。指定中文后，英文句子依然写成英文。在 50 条“一句中文接一句英文”的短录音里，第二句全部保留，没有一句被翻译。Friday never leaves the language on auto-detect: when Qwen3-ASR guessed the language itself, it translated Chinese inside mixed dictation into English. Set to Chinese, it still writes English sentences in English. On 50 short "one sentence Chinese, one sentence English" recordings, every second sentence survived and none was translated.

只是一个人、一个安静房间里的结果，不是通用评测。One speaker, one quiet room. Not a general benchmark.

松手到出字约 0.6 秒About 0.6 s from release to text

在作者的 41 条测试录音上，换成 Qwen3-ASR 的 0.5.0 中位延迟约 0.6 秒（0.4.0 的 Whisper Medium 约 1.6 秒）；中文错字率 1.2%，英文错词率 1.3%。不靠网络，不受网速影响。On the maintainer's 41 evaluation recordings, 0.5.0 (Qwen3-ASR) has a median latency of about 0.6 s (vs ~1.6 s for Whisper Medium in 0.4.0), with 1.2% Chinese character error rate and 1.3% English word error rate. No network involved.

同样只代表作者一个人的声音。Again, from one speaker.

完全本地，模型逐个文件校验Fully on-device, model files verified

录音只在你的 Mac 上转写，不上传、不远程记录。模型从固定版本下载，每个文件都做 SHA-256 校验，对不上就拒绝。Audio is transcribed on your Mac and never uploaded or logged remotely. The model comes from a pinned revision and every file is checked against a SHA-256; anything that doesn't match is rejected.

不是独有：Handy、VoiceInk、TypeAloud 也在本地识别。Not unique: Handy, VoiceInk, and TypeAloud also transcribe locally.

免费，MIT 开源Free and MIT licensed

不用订阅，不要账号。代码在 GitHub，随便看、随便改。No subscription, no account. The code is on GitHub to read and change.

不是独有：Handy（MIT）、TypeAloud（GPL-3.0）也免费开源。Not unique: Handy (MIT) and TypeAloud (GPL-3.0) are free and open source too.

缺点写在明处Limitations stated up front

没测过别人的声音、句末不加标点、会改写像 AI 工具名的词，全都写在下面和仓库的 LIMITATIONS 里。Untested on other voices, no sentence-final punctuation, opinionated rewrites of AI tool names: all listed below and in the repo's LIMITATIONS.

日常用得顺手Small things done right

  • 快捷键可选右 Command（默认）、左 Command、左右 Option、右 ControlPush-to-talk on Right Command (default), Left Command, Left or Right Option, or Right Control
  • 文字插入光标处，之后还原原来的剪贴板Pastes at your cursor and restores your clipboard
  • 遇到密码框等受保护输入，拒绝粘贴Refuses to paste into password fields
  • 首次引导：开权限、下模型First-run onboarding for permissions and the model

安装Install

运行要求Requirements

  • Apple 芯片的 Mac，macOS 14 或更新Apple Silicon Mac, macOS 14+
  • 运行时约占 2.3 GB 内存About 2.3 GB of free memory while Friday runs
  • 第一次设置时会下载约 2.2 GB 的识别模型About 2.2 GB for the speech model, downloaded during first-run setup

步骤Steps

  1. 从 Releases 下载最新的 zip，把 Friday.app 拖进“应用程序”。Download the latest zip from Releases and move Friday.app to Applications.
  2. 现在的安装包还没签名，第一次打开会提示“Apple 无法验证 Friday”，只给“完成”和“移到废纸篓”两个按钮。点 完成。Builds are not signed yet, so the first launch says "Apple could not verify Friday" and only offers Done or Move to Trash. Click Done.
  3. 打开 系统设置 → 隐私与安全性。Open System Settings → Privacy & Security.
  4. 在 安全性 里，点 Friday 旁边的 仍要打开。Under Security, click Open Anyway for Friday.
  5. 再点一次 仍要打开，输入电脑密码。Confirm Open Anyway again and enter your administrator password.

然后按提示给三个权限，这是 macOS 对语音输入工具的要求：Then grant the three permissions macOS requires for a voice-input tool:

  1. 麦克风：听你说话Microphone to hear you
  2. 辅助功能：把字贴进别的 AppAccessibility to paste into other apps
  3. 输入监控：识别你按下的快捷键Input Monitoring to detect the push-to-talk key

最后下载 Qwen3-ASR 1.7B 模型，选好快捷键，按住说话，松手，就可以了。Install the Qwen3-ASR 1.7B model, pick a key, hold it, speak, release.

和类似工具对比How it compares

只写我们核实过的；没核实的标“未确认”。各项目可能已更新，以它们自己的页面为准。Only facts we checked; anything unchecked is marked "unverified". Projects change, so their own pages are the source of truth.

工具Tool 识别在哪跑Recognition 中英混说Mixed zh/en 许可证License 价格Price 平台Platforms
Friday本机On device主打；语言总是指定，公开 50 条测试Core focus; language always set, 50-clip test publishedMIT免费FreeApple 芯片 MacApple Silicon Mac
TypeAloud本机（Qwen3-ASR）On device (Qwen3-ASR)主打；模型从音频自动判断语言；有词典和自动替换Core focus; model detects language from audio; has dictionary and auto-replacementsGPL-3.0免费（可选云端转写用你自己的 Groq / OpenAI key，费用由服务商收）Free (optional cloud transcription with your own Groq / OpenAI key; provider charges apply)Mac（主力）；Windows、Linux 实验版Mac (primary); Windows, Linux experimental
Handy本机On device没有专门调过Not specifically tunedMIT免费FreeMac / Windows / Linux
VoiceInk本机On device没有专门调过Not specifically tunedGPL-3.0源码免费，官方构建版付费Source free; official builds paidMac
CapsWriter-Offline本机（也用 Qwen3-ASR）On device (also Qwen3-ASR)未确认UnverifiedMIT免费FreeWindows
Wispr Flow云端Cloud未确认Unverified闭源Proprietary免费版（桌面每周 2,000 词）；Pro $15/月或 $144/年Free tier (2,000 words/week on desktop); Pro $15/mo or $144/yrMac / Windows / iOS / Android

如果你主要说英文，或者需要 Windows、Linux，Handy 更合适。需要自定义词表的话，TypeAloud 现在就有，Friday 还没有。If you mostly dictate English or need Windows or Linux, Handy is a better fit. If you need a custom vocabulary today, TypeAloud has one and Friday doesn't yet.

常见问题FAQ

我的录音会上传吗？Is my audio uploaded?

不会。录音只在你的 Mac 上转写，从不上传、不远程记录、不发给第三方。没有账号，没有统计。只有第一次下载模型文件时会联网，直接从源头下载。No. Audio is transcribed on your Mac and never uploaded, logged remotely, or sent to third parties. No account, no telemetry. The only network use is downloading the model files the first time, straight from their source.

能断网用吗？Does it work offline?

可以。模型下载好之后，识别全部在本机完成，不需要网络。Yes. Once the model is downloaded, recognition runs entirely on your Mac with no network needed.

准不准？How accurate is it?

在作者自己的 41 条录音上，中文错字率 1.2%，英文错词率 1.3%。这只是一个人、一个安静房间的结果，你的口音和环境可能不一样。欢迎把你的结果发到 Issues。On the maintainer's 41 recordings: 1.2% Chinese character error rate, 1.3% English word error rate. That's one speaker in one quiet room; your accent and room may differ. Results posted to Issues are welcome.

一句话里中英混着说真的行吗？Does mixing Chinese and English in one sentence really work?

这是 Friday 存在的理由。识别语言总是指定（默认中文，也可选英文），英文句子依然写成英文。50 条“一句中文接一句英文”的录音里没有一句被翻译或丢掉。注意：像 cloud、codec 这类词会被改写成 Claude、Codex。It's why Friday exists. The language is always set (Chinese unless you choose English), and English sentences stay English. In 50 "one sentence Chinese, one sentence English" recordings, none was translated or dropped. Note that words like "cloud" and "codec" get rewritten to "Claude" and "Codex".

要钱吗？What does it cost?

免费，MIT 开源，没有订阅。Nothing. Free and MIT licensed, no subscription.

为什么第一次打开提示“无法验证”？Why does macOS say it can't verify Friday?

目前的安装包还没有签名。按安装步骤在“隐私与安全性”里点“仍要打开”即可。Current builds aren't signed yet. Follow the install steps and click Open Anyway in Privacy & Security.