Run Whisper on your iPhone
Your voice gets turned into text on the phone. No upload, no account, works in airplane mode. Here is what actually does that today.
Whisper is OpenAI’s open speech-to-text model. The Whisper page has every size. This guide is only about the phone. I have not timed any of this on a phone, so there is no phone speed marked Measured. Where someone else measured, I say who.
The easy way: an app
Each of these says on its own App Store page or site that it transcribes on the device. I checked the listings on 19 Nov 2026. Prices and versions change, so I am not copying them here.
- Aiko (Sindre Sorhus). Built on Whisper and whisper.cpp. The App Store listing says it uses “the Whisper medium or small model on iOS depending on available memory” and “Nothing leaves your device.” One-time purchase. App Store · sindresorhus.com/aiko
- Whisper Notes (SmashMelon). The listing says “100% offline on your iPhone and iPad” and lets you pick Whisper, Parakeet or SenseVoice. It says it needs an iPhone 12 or newer. One-time purchase. App Store · whispernotes.app
- Whisper Transcription (Jordi Bruin, the MacWhisper developer). It has both: a cloud mode and a downloadable on-device model for offline use. The listing says on-device needs an iPhone 13 or newer, or an M-series iPad. Make sure you pick the on-device model. App Store · macwhisper.com
- Superwhisper. A dictation keyboard with cloud and local models. The listing says you can “Run local models right on your device.” Pick a local model if you want it offline. App Store · superwhisper.com
These are the makers’ listings. None of them is a shop link.
Which sizes fit which chip
Vendor-reported This comes from Argmax, the company behind WhisperKit. Their public device support list names Apple chips and model identifiers. It does not name iPhone models. It is their list for their software, not a guarantee for every app.
| Chip Argmax lists | Sizes Argmax lists | WhisperKit default |
|---|---|---|
| A12, A13 | tiny, base (and .en) | tiny |
| A14 | tiny, base, small (and .en) | base |
| A15 and newer | tiny, base, small, plus compressed large-v2, large-v3 and large-v3-turbo builds (about 600–950 MB) | base |
Apple’s own spec pages name the chip in a phone. These are examples, not a full catalog:iPhone XS, A12 Bionic,iPhone 12, A14 Bionic,iPhone 13, A15 Bionic,iPhone 14, A15 Bionic,iPhone 15, A16 Bionic. I am not filling in phones whose spec page I have not opened.
Medium is not on Argmax’s list for any iPhone. Full-size large builds are listed only for M-series iPads and Macs. Argmax recommends large-v3-v20240930_626MB (a compressed large-v3-turbo) “across iOS and macOS for maximum accuracy.”
Not everyone agrees. Aiko’s FAQ says “Even the latest iPhone is not powerful enough to run the large model.” That is the Aiko developer talking about his app, which uses whisper.cpp. Both are vendor claims about their own software.
What other people ran into
Community Single reports, not tests I repeated.
- The developer of Koedori wrote on 17 Nov 2026 that on their iPhone 14 the first load of large-v3-turbo through WhisperKit took 197 to 205 seconds while the Neural Engine compiled it. A reload took 4 seconds. They switched their app to Apple’s built-in SpeechTranscriber.
- The WhisperMetalKit author says large-v3-turbo would not compile through Core ML on their iPhone, but ran through whisper.cpp’s Metal backend.
- A whisper.cpp issue from 2024 reports turbo crashing on an iPhone 12 mini’s GPU after a few segments.
The takeaway: the first run of a big model on a phone can take minutes. Start with base or small. Try turbo after that works.
For tinkerers: build it yourself
- whisper.cpp has iOS sample apps, whisper.objc and whisper.swiftui. The README shows a demo “running the model on an iPhone 13 device - fully offline, on-device.” You need a Mac with Xcode. On Apple Silicon the encoder can run on the Neural Engine through Core ML; the README says that can be “more than x3 faster compared with CPU-only.” That speed is Vendor-reported.
- WhisperKit by Argmax is a Swift package, now inside argmax-oss-swift (MIT). It downloads the recommended model for the device on its own. Prerequisites listed: macOS 14 and Xcode 16.
What about Wispr Flow?
Wispr Flow is a popular voice keyboard for iPhone. It is cloud-only. Its Data Controls page says “Transcription always occurs on the cloud.” Its help center says the iPhone keyboard “needs Wi-Fi or cellular data to transcribe.” Since 17 Sep 2026 it runs on Wispr’s own model, Canto, not on Whisper (what’s new). Notetaker on iPhone, announced 18 Nov 2026, keeps recording when the signal drops and “processes the transcript and summary once the phone reconnects” (Wispr’s post). The App Store listing is Wispr Flow. It is not local AI, so it does not get a model page.
The one already on your phone
iOS has its own on-device speech recognition. Apple’s developer framework is SpeechAnalyzer. Aiko’s FAQ says Apple’s transcriber “supports far fewer languages” than Whisper and has no automatic language detection. If it does what you need, it is already there.