Talk into any text field
Keep the audio on your Mac
任何输入框都能说话
音频留在你的 Mac 上

Double-tap Control, speak, and the text lands where your cursor already is
Two of the three engines never send audio anywhere
双击 Control 说话,文字落在光标所在的位置
三个引擎里有两个,音频不出这台 Mac

Apple Silicon · Requires macOS 26 or later · 5-day free trial · 7-day refund · 31 MB Apple Silicon · 需要 macOS 26 或更新版本 · 免费试用 5 天 · 7 天退款 · 31 MB

Local doesn't mean second best 本地不等于将就

Same 10-minute conference recording. Eight engines, nine runs, one script. 同一段 10 分钟会议录音,八个引擎九路对照,一个脚本。

Runs on跑在哪 Developer terms caught行话命中
GPT-4o Transcribe Cloud云端 36
AssemblyAI Universal-3.5 Pro Cloud云端 36
Whisper large-v3-turbo Your Mac你的 Mac 35
Deepgram Nova-3 Cloud云端 35
Speechmatics Melia 1 Cloud云端 18

That one-point gap is smaller than the gap between two runs of the same engine. We paid for five cloud APIs to check our own. 这一分的差距,比同一家引擎自己跑两遍的差距还小。这五家云端 API 是我们自己付费买来核自己的。
Full test: all nine runs, method, and what went wrong on the way → 完整实测:九路对照、测法与翻车记录 →

Three engines, three jobs. 三个引擎,三种活

Fastest to first word: Apple Speech. Throughput: Whisper tiny. Accuracy: Whisper large. Apple Speech 出字最快,Whisper tiny 跑得最快,Whisper large 最准。

Where audio goes音频去哪 Languages语言 Setup配置
Apple Speech Stays on device不出设备 System languages, auto-detect系统语言,自动识别 None
Local Whisper Stays on device不出设备 Multilingual多语言 One in-app download应用内下载一次
Cloud To your provider直连你的厂商 Provider-dependent取决于厂商 Your own key. Off by default你自己的 key,默认关闭
First word出第一个字 Speed速度
Apple Speech 292 ms 14× real time14 倍速
Whisper tiny 349 ms 48× real time48 倍速
Whisper large-v3-turbo 912 ms 8× real time8 倍速
One hotkey, everywhere一个快捷键,处处可用 The mic lives on the system, not inside one app. It works in the terminal, where agents live and nothing else listens.麦长在系统上,不长在某个 app 里。终端里也能用,那是 agent 住的地方。
The pill follows you胶囊跟着你走 A small capsule shows what it heard, across full-screen Spaces. Stop speaking and the text lands.一枚小胶囊实时显示听到的话,跨全屏 Space 也在。说完即上屏。
31 MB, one purchase31 MB,一次买断 The whole app is a 31 MB download, dictation included. You buy it once.整个应用 31 MB,听写内含。一次买断。

Dictation is one thing Puffy does.
Any link in, clean markdown out. Web pages, videos, PDFs.
16x fewer tokens than handing your agent the raw page.
听写只是 Puffy 做的其中一件事。
任何链接进,干净的 markdown 出。网页、视频、PDF 都一样。
比把原始页面直接甩给 agent 少 16 倍 token。

See everything Puffy does 看看 Puffy 的全部 The token test → token 实测 →
Activate once, then use it offline. 激活一次,之后离线可用。