A speech AI buffet for Apple Silicon, no cloud required

What it does

speech-swift is a modular Swift package that wraps two dozen speech and audio models—ASR, TTS, diarization, noise suppression, even text-to-music—into Apple-native backends. Everything runs locally on Apple Silicon via MLX or CoreML, so no audio leaves the device. You import only the SPM modules you need, from an 82M-parameter TTS engine up to a 7B speech-to-speech model.

The interesting bit

The breadth is almost comical: transcription across 1,672 languages, 90-minute podcast generation, real-time noise suppression, and full-duplex speech-to-speech conversation, all quantized down to INT4 for laptop memory. It functions like a native-Swift porting house for NVIDIA, Meta, and Alibaba research checkpoints, repackaging them into Apple-friendly binaries.

Key highlights

• Modular SPM targets mean you pay compile and memory cost only for the models you actually ship.
• Backends are split between MLX (flexible research models) and CoreML / Neural Engine (efficient production inference).
• Standout models include Omnilingual ASR (1,672 languages), PersonaPlex (7B full-duplex speech-to-speech), and Kokoro TTS (82M, iOS-ready).
• Includes oddball audio utilities like source separation, 48 kHz super-resolution, and MAGNeT text-to-music generation.
• A companion SpeechUI module offers a ready-made SwiftUI transcription view.

Verdict

Ideal for Swift developers building privacy-first voice agents or transcription tools on macOS and iOS. If you’re not targeting Apple Silicon or you need cloud-scale batching, this is the wrong stack.