WhisperKit is an on-device inference framework, a library that runs a trained AI model on the user's own device instead of on a server. Its maker, Argmax, built it to run OpenAI's Whisper speech-to-text models on Apple Silicon. Argmax now ships it as one library inside the Argmax Open-Source SDK (repository argmax-oss-swift), next to SpeakerKit for speaker diarization (labelling who speaks when) and TTSKit for text-to-speech.
A developer adds the Swift package in Xcode, creates a WhisperKit object and calls transcribe with the path of an audio file (WAV, MP3, M4A or FLAC); the result is the text. Unless the app names a model, WhisperKit downloads the recommended Whisper model for the device, from tiny to large-v3 turbo. Long recordings can be read from disk in chunks split at silences. A command line app installs through Homebrew and includes a local server that implements the OpenAI Audio API, so existing OpenAI SDK clients can use it.
The SDK is free and open source under the MIT license and needs macOS 14 and Xcode 16 or later. Argmax sells a separate Pro SDK with more models and real-time transcription.