WhisperKit is an open-source Swift framework from Argmax that runs OpenAI's Whisper speech recognition models on Apple devices, using Apple's Core ML. App developers build it into Mac and iPhone apps so speech can be turned into text on the device; most people meet it inside an app rather than installing it. It is released under the MIT licence, and its code now lives in Argmax's argmax-oss-swift repository on GitHub, where the old WhisperKit address redirects.
Publisher note: SuprFlow publishes this guide and is one of the products discussed. Competitor details come from each company's own pages, checked on 1 October 2026. Prices and features change; confirm before you buy.
What exactly is WhisperKit?
The argmax-oss-swift README describes the package as "a collection of turn-key on-device inference frameworks", with three parts:
- WhisperKit, "for speech-to-text with OpenAI Whisper"
- SpeakerKit, for labelling who spoke when
- TTSKit, for text-to-speech
Key facts from the README and licence file:
| WhisperKit | |
|---|---|
| Maker | Argmax, Inc. |
| Licence | MIT ("Copyright (c) 2024 argmax, inc.") |
| Language | Swift, added to an app with Swift Package Manager |
| Prerequisites | macOS 14.0 or later and Xcode 16.0 or later, to build with it |
| Models | Whisper in Core ML format, hosted on Hugging Face; it "automatically downloads the recommended model for the device if not specified" |
| Command line | brew install whisperkit-cli |
| Local server | Implements the OpenAI Audio API, so existing OpenAI client code can talk to a model on your own machine |
Argmax also sells a Pro SDK. The README lists features such as "Real-time transcription with speakers" and "custom vocabulary" there, so those are not part of the free package.
How does WhisperKit relate to OpenAI's Whisper?
Whisper is the model; WhisperKit is one way to run it. OpenAI's Whisper repository describes Whisper as "a general-purpose speech recognition model" that can do multilingual speech recognition, speech translation and language identification, and says its "code and model weights are released under the MIT License".
OpenAI's own code is in Python. WhisperKit takes the same model family, converts it to Core ML, and runs it from Swift. Argmax publishes the converted models in its Hugging Face repository and a companion tool, whisperkittools, for converting your own fine-tuned Whisper models.
How is it different from whisper.cpp?
whisper.cpp is the other popular way to run Whisper on a Mac. Its README calls it "High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model" in "Plain C/C++".
| WhisperKit | whisper.cpp | |
|---|---|---|
| Maker | Argmax | ggml-org |
| Language | Swift | C/C++ |
| Licence | MIT | MIT |
| How it uses Apple hardware | Runs Core ML versions of Whisper | "On Apple Silicon, the inference runs fully on the GPU via Metal"; optional Core ML lets the encoder run on the Apple Neural Engine |
| Platforms named in the README | Apple platforms (macOS 14 or later to build) | Mac (Intel and Arm), iOS, Android, Linux, Windows and more |
| Typical user | A developer building a Swift app | A command-line user or a cross-platform project |
Neither is "better" in general. Whisper vs faster-whisper vs WhisperX compares all five common Whisper runtimes.
What do Core ML and Apple Silicon support mean?
Core ML is Apple's framework for running machine-learning models inside apps. Apple's Core ML documentation says it "optimizes on-device performance by leveraging the CPU, GPU, and Neural Engine while minimizing its memory footprint and power consumption", and that "Running a model strictly on a person's device removes any need for a network connection".
In plain terms: a Whisper model converted to Core ML can use the processor, the graphics chip and the Neural Engine in an Apple Silicon Mac, as Core ML decides. The WhisperKit README does not promise which chip handles each step, so treat claims about "runs on the Neural Engine" as the app maker's, not WhisperKit's.
Two practical points follow:
- On-device recognition is not the same as no network. WhisperKit downloads models from Hugging Face, so a first run needs a connection, and the app around it may use the network for other things. On-device speech recognition explained covers this.
- The model size is a separate choice. The README lists Tiny, Base, Small and Large v3 Turbo variants, several with English-only versions. Argmax recommends a compressed Large v3 Turbo for "maximum multilingual accuracy" and Tiny only for development and debugging. Whisper model sizes explains the trade-offs.
Which apps use WhisperKit?
Only trust an app's own statement. Many Mac apps say they use "Whisper" without naming the runtime, and "Whisper" alone could mean WhisperKit, whisper.cpp or OpenAI's cloud API. BetterDictation's home page, for example, says it is "Powered by OpenAI's Whisper, running entirely on Apple's Neural Engine" without naming a framework (BetterDictation). If the runtime matters to you, ask the developer.
SuprFlow 1.0.0, which publishes this guide, recognises speech on the Mac with a WhisperKit (Whisper) model. It offers Small and Medium models, each in multilingual and English-only versions, and asks for a speech-recognition language during setup. It requires an Apple Silicon Mac with macOS 14 or later. Using WhisperKit is an implementation detail, not a guarantee of accuracy; we have not benchmarked SuprFlow against other apps. Compare SuprFlow speech models shows how to choose between its models.
Should you use WhisperKit directly?
- You are building a Swift app for Mac or iPhone: yes, it is designed for that, and the MIT licence allows commercial use under its terms.
- You want to transcribe files from Terminal:
whisperkit-clior whisper.cpp both work. How to run Whisper on a Mac walks through the options. - You just want to dictate: use Apple Dictation or a dictation app. A framework has no hotkey, no text insertion and no settings screen.
Questions people ask
Is WhisperKit made by OpenAI?
No. OpenAI made the Whisper models. WhisperKit is made by Argmax, Inc., which converts and runs those models on Apple devices.
Is WhisperKit free?
The open-source package is free under the MIT licence. Argmax sells a separate Pro SDK with extra features such as real-time transcription with speakers and custom vocabulary.
Does WhisperKit work offline?
The recognition runs on the device once a model is downloaded. The first model download needs a connection, and any app built on WhisperKit may use the network for other things.
Is WhisperKit better than whisper.cpp?
They run the same model family in different ways. WhisperKit suits Swift apps on Apple platforms; whisper.cpp suits command-line and cross-platform use. We have not run a controlled benchmark of the two.
Related guides
- Whisper vs faster-whisper vs WhisperX
- On-device speech recognition explained
- How to run Whisper on a Mac
- Whisper model sizes
- Parakeet vs Whisper speech models
- Compare SuprFlow speech models