Speech Swift is an open-source AI speech toolkit for Apple Silicon, published on GitHub by Soniqo, that runs speech recognition, text-to-speech, speech-to-speech translation, voice activity detection and speaker diarization entirely on-device with MLX and CoreML. The project has gathered around 1,200 GitHub stars and is actively developed in 2026.
Speech Swift ports dozens of leading speech models to native Swift for Mac and iOS: Qwen3-ASR, Whisper, Parakeet and Nemotron for transcription in up to 1,600+ languages via Meta’s Omnilingual ASR; Qwen3-TTS, CosyVoice, Kokoro and IndexTTS2 for synthesis and voice cloning; plus forced alignment, speech enhancement, source separation, wake-word detection and full voice pipelines (ASR → LLM → TTS) for voice agents. Everything runs locally on the Neural Engine and GPU — no cloud, no API keys, no data leaving the device — with published on-device benchmarks down to iPhone-class hardware. It installs via Homebrew or Swift Package Manager, ships demo apps and an optional HTTP API server, and is already embedded in a growing list of Mac and iOS apps for dictation, transcription and voice interaction.
Speech Swift is free and open source under the Apache 2.0 license. Some bundled models carry their own licenses (a few are research/non-commercial), noted in the documentation.
iOS and macOS developers who want private, offline speech intelligence in their apps without paying per-minute cloud transcription fees.