WhisperKit

WhisperKit compiles OpenAI's Whisper models to Core ML so transcription runs entirely on Apple devices, using the Neural Engine on Apple silicon with no server or network round trip. The Swift package delivers streaming transcription with realtime partial results, word and segment timestamps, voice activity detection for chunking, automatic language detection, and memory-efficient incremental loading for long recordings. Model options span tiny through large-v3 turbo, including a compressed 626 MB turbo build that keeps multilingual accuracy practical on phone-class hardware, and custom fine-tuned models can be pulled through Hugging Face. Developers add it via Swift Package Manager on macOS 14 or newer, a command line build installs with brew install whisperkit-cli, and a bundled local server exposes an OpenAI-compatible audio API with server-sent event streaming. Argmax ships it under the MIT license at around 6,300 stars, and since version 1.0 the repository has been renamed argmax-oss-swift, pairing WhisperKit with SpeakerKit diarization on Pyannote and TTSKit on Qwen3-TTS, alongside a commercial Pro SDK.