Hold a key, speak, and the words appear at your cursor. Transcribed entirely on your Mac. No cloud, no account, no subscription.
Buy once — $29Transcription runs on your Mac's own silicon. Your voice never leaves your machine: there is no server to send it to.
Hold the key, speak, release. The words are typed at your cursor: mail, notes, code, chat, anywhere text goes.
No subscription, no account, no free tier holding features hostage. Buy it, own it.
Two honest exceptions, because you'd find them anyway:
1. The model download. On first launch the app downloads its speech model (~626 MB, from Hugging Face). Once. After that, the only traffic you'll ever see is localhost, and only if you enable the Ollama extras below.
2. Failed transcriptions are kept, on purpose. Your voice is written to a temporary file on your Mac, transcribed on your Mac, and deleted the moment it's typed. If transcription fails, we keep that file, at~/Library/Application Support/SoundPoof/failed-audio/, rather than silently losing your words. It sits on your disk. Delete it whenever you want. It still never leaves your machine, because there is nowhere for it to go.
Out of the box SoundPoof does raw dictation with zero setup. If you already run Ollama, you can switch on optional cleanup (fixes filler words and punctuation) and voice commands. Everything stays on your machine; Ollama is localhost by definition.
Every dictation you speak into Wispr Flow goes to their servers, by architecture. Their own FAQ: "Wispr's backend must decrypt audio to perform transcription."Privacy Mode limits retention, not cloud processing. In 2025 they shipped a screenshot feature that sent screen contents to their servers; the CTO apologized, and Screen OCR is now off by default. Wispr Pro is $144/yr, and when their servers have a bad day, so does your dictation (their own help center: "Transcription suddenly got worse or feels less accurate").
SoundPoof can't have any of these problems. Not because we promise better, but because there is no us. Your data never exists anywhere to protect.
# sources: wisprflow.ai FAQ & pricing · Wispr help center · verified 2026-07-13
Built by a nurse. I spent years around information that legally and morally must not leak. I wanted dictation I could trust in the same way: not a privacy policy, an architecture.
SoundPoof isn't HIPAA-certified anything and doesn't claim to be. It's simpler than that. Audio is captured, transcribed on-device, typed at your cursor, and deleted. The one time it isn't deleted, a transcription that failed, it stays on your disk and nowhere else.
They're good projects. If you like building from source and tinkering, use them, genuinely. SoundPoof is for people who want the boring version: download a signed, notarized app, hold a key, and it works. Silence handling, punctuation, a first-run permissions flow that makes sense, and one person accountable when something breaks. $29 buys the part that open source gives away last: it's finished.
Use it if it works for you. SoundPoof exists because Whisper-class models are a generation ahead on accuracy (technical vocabulary, punctuation, fast messy speech) and because hold-to-talk beats toggle-and-hope: press, speak, release, done, in any app, with no listening state left behind.
The transcription engine is WhisperKit, credited and paid for in effort like everyone else's. What you're buying is everything around it: instant hold-key capture, typing at your cursor in any app, the permissions flow, silence detection, optional local cleanup, and the guarantee that the whole pipeline never touches a server. "Wrapper" is the compliment; the wrapping is the product.
Poof as in vanish: the sound of your audio ceasing to exist, deleted the moment it's typed. (The exception is a transcription that fails, where we keep the recording on your disk so you don't lose your words. See the two honest exceptions above.) No slight intended in any dialect.