Talk instead of typing, in any app. Speech recognition runs on your device, transcripts live in your own iCloud, and there is no account and no server.
Download for MacDownload for iOS
The Mac and iPhone apps are not released yet. Want to try them before launch? Request early access and tell us which platform you use; we will write to you when a test build is ready.
Hold a hotkey, speak, let go. The text appears in the app you are using: mail, chat, notes, your code editor. Toggle mode for longer passages.
Switch to the SayPrivate keyboard in any app, tap the mic and speak. The text is inserted where your cursor is.
Record a meeting or a voice note and get a transcript with timestamps. On Mac it can capture your microphone and the call audio together.
Every dictation and recording is saved, searchable and synced between your Mac and iPhone through your own iCloud.
Export any transcript as Markdown, plain text, SRT subtitles or JSON. No lock-in.
Delete one item, everything on this device, or everything in iCloud, from Settings in the app.
Cloud dictation apps send your voice to their servers to turn it into text. That means your drafts, your meetings and anything you say near the mic pass through someone else's computers. SayPrivate does the same job with on-device speech recognition built into macOS and iOS, so your audio never leaves your device.
| SayPrivate | Cloud dictation apps | |
|---|---|---|
| Where speech is recognized | On your device | On the vendor's servers |
| Where fillers and corrections are cleaned up | On your device | On the vendor's servers |
| Where transcripts are stored | Your own iCloud private database | The vendor's cloud |
| Account required | No | Usually |
| Works offline | Yes | No |
| Export and delete everything | Anytime, in the app | Varies |
| Speech model | NVIDIA Parakeet TDT 0.6B v3, run through FluidAudio on the Neural Engine of your iPhone or Mac. It covers 25 European languages, including English, German, French, Spanish, Russian and Ukrainian, and on the Open ASR Leaderboard averages about 6.3% word error rate in English, against 7.8% for Whisper large-v3-turbo. Punctuation and capitals come from the model. For other languages, and until Parakeet is downloaded, SayPrivate falls back to Apple SpeechAnalyzer, the speech recognition built into iOS 26 and macOS 26. |
| Model download | The Parakeet model (about 480 MB) downloads once from Hugging Face; Apple's speech models download once from Apple. You can delete either in Settings > Models. After that, dictation works in airplane mode. |
| Punctuation and cleanup | Apple's on-device language model (Foundation Models framework) turns the raw transcript into the text you meant: it drops filler words and false starts, applies your corrections ("no, I mean..."), adds punctuation and lists, and keeps the words from your personal dictionary spelled your way. On Mac it matches the app you dictate into: short for chat, full sentences for email, exact terms for code, lists for notes. If the model is off or slow, local rules clean the text instead. Optionally you can download Qwen3 1.7B (about 1 GB, run with MLX) as the cleanup model. None of it calls a server. |
| Mac text insertion | macOS Accessibility inserts the text at the cursor. If an app does not accept it, SayPrivate pastes with Cmd+V and restores your clipboard. |
| Mac meetings | Microphone plus system audio through a Core Audio process tap (macOS asks for permission), mixed into one file on your Mac and transcribed locally with timestamps. |
| iPhone keyboard | iOS does not let keyboards use the microphone, so the keyboard asks the SayPrivate app to listen and receives the text through a shared App Group container on the device. The keyboard makes no network requests. |
| Storage and sync | SwiftData on the device, synced through the CloudKit private database of your Apple Account. No SayPrivate server exists. |
| Network requests | iCloud sync (Apple), one-time model downloads (Hugging Face for Parakeet and the optional Qwen3 model, Apple for its speech models), subscription handling on iPhone (Apple, Superwall) and in the Mac App Store version (Apple), license check in the direct Mac version (Dodo Payments or Gumroad). Audio and transcripts are never part of them. |
| Requirements | iPhone 15 Pro or newer with iOS 26. Mac with Apple M1 or newer with macOS 26. Older devices lack the Neural Engine and on-device language model the app is built on. |
| Credits | Parakeet TDT 0.6B v3 by NVIDIA, licensed under CC BY 4.0, Core ML conversion by FluidInference. FluidAudio (Apache 2.0). Qwen3 1.7B by the Qwen team (Apache 2.0). MLX Swift (MIT). Illustrations by Icons8. |
iPhone subscriptions renew automatically unless turned off at least 24 hours before the end of the current period. Manage or cancel them in Settings > [your name] > Subscriptions. See the Terms of Use.
Yes. Speech recognition and text cleanup run on your device. The speech model is downloaded once (from Hugging Face, or from Apple for Apple's fallback model); after that, dictation works in airplane mode. The only network use is that one-time download, optional iCloud sync to your own account, the purchase and license check, and on iPhone the subscription paywall.
On your device and, if iCloud is on, in your own iCloud private database (Apple CloudKit) so your Mac and iPhone stay in sync. We have no server and no copy. With iCloud off, nothing leaves the device.
In the app, open Settings and choose to delete a single item, everything on this device, or everything in iCloud. Deleting the app removes its local data; iCloud data can also be removed in Settings > [your name] > iCloud > Manage Account Storage.
The SayPrivate keyboard only inserts the text you dictate. It does not record, store or send your keystrokes and makes no network connections. It asks for Full Access only so it can receive the dictated text from the SayPrivate app through a shared container on your iPhone.
Parakeet covers 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. For other languages SayPrivate uses the languages Apple's on-device speech recognition supports, such as Chinese, Japanese and Korean, and downloads that model once.
Both. The Mac App Store version shares your iPhone subscription and inserts text into other apps only after you turn on Direct Insert, which asks for Accessibility access; until then it copies the text for you to paste with Command-V. The direct version from this site is a license and inserts text out of the box. Both run the same on-device models and sync with your iPhone through your iCloud.
iPhone: an iPhone 15 Pro or newer with iOS 26. Mac: a Mac with Apple M1 or newer with macOS 26. Earlier iPhones, including iPhone 15 and 15 Plus, and Intel Macs lack the Neural Engine and on-device language model SayPrivate is built on. Apple Intelligence should be turned on for the best cleanup; without it SayPrivate uses simpler local rules. On Mac, SayPrivate asks for Microphone access, Accessibility access to type into other apps (in the Mac App Store version only after you turn on Direct Insert), and audio capture permission to record call audio in meeting mode.
SayPrivate is in development. Request early access for Mac, iPhone or both.