Speech Swift
An open-source collection of on-device speech AI models for Apple Silicon, covering speech recognition, synthesis, speech-to-speech, voice activity detection, and speaker diarization.
Speech Swift is a collection of AI speech models for Apple Silicon, built with MLX Swift and CoreML, that run entirely on-device on Mac and iOS for speech recognition, synthesis, and understanding, with no cloud calls, API keys, or data leaving the device.
It packages many published speech models behind a common interface, grouped into capability areas such as speech-to-text/ASR (including Qwen3-ASR, Whisper, Parakeet, Omnilingual ASR, and low-latency streaming models), alignment, text-to-speech, speech-to-speech, enhancement and restoration, source separation, music and audio generation, and wake word/VAD/diarization and speaker identity. Models run through the Apple Neural Engine or MLX, with reported real-time factors as low as roughly 0.04 to 0.28 on recent iPhone and Mac hardware.
It is used as a dependency by other public speech and voice applications, and is documented in more than a dozen languages.