**Speech Swift** packages a large catalog of speech models as native Swift libraries for Mac and iOS, running entirely on-device through **MLX Swift and CoreML** — no cloud, no API keys, no data leaving the machine. It is Apache-2.0, published by the **soniqo** organization, and has reached **1,139 stars and 151 forks** since going public on **4 February 2026**. Releases ship at a rapid clip, with **v0.0.26** on **17 August 2026** following two more in the preceding day. The transcription lineup is the strongest part of the catalog. It spans **Qwen3-ASR** (52 languages, MLX and CoreML), **Whisper Large-v3 Turbo** through a native CoreML runtime on the Neural Engine, **Parakeet TDT** (NVIDIA FastConformer with a TDT decoder, 25 languages), **Cohere Transcribe 2B**, **Voxtral Mini 3B**, and Meta's **Omnilingual ASR** covering **1,672 languages across 32 scripts**. **MOSS Transcribe Diarize** does offline transcription with model-generated speaker labels and timestamps at a 128K MLX context. Two NVIDIA Nemotron streaming variants handle low-latency ASR with native punctuation and capitalization — a multilingual build at 40 language-locales and a smaller English-only one. The project publishes real-time-factor figures per model rather than a single headline number: Omnilingual ASR at 0.28 RTF on an iPhone 16 Pro, Voxtral Mini at 0.074 RTF on an M5 Pro, and streaming dictation at 0.04 RTF on an iPhone 16 Pro. Coverage extends well past recognition. TTS includes **Qwen3-TTS, CosyVoice, VoxCPM2, IndexTTS2, F5-TTS, Higgs TTS 3, Kokoro, VibeVoice, Magpie, Supertonic, Chatterbox, OmniVoice, Indic-Mio, and Sesame CSM**. There are speech-to-speech and voice-agent paths (**Hibiki Zero-3B** streaming translation, **PersonaPlex 7B** full-duplex, **VoiceChat 11B**), on-device LLM and translation modules (Qwen3Chat, FunctionGemma for tool calls, MADLAD-400 across 400+ languages), enhancement and separation (**DeepFilterNet3**, **LocalVQE AEC**, HTDemucs and Open-Unmix source separation, **FlashSR** super-resolution), music generation via **MAGNeT** and **Stable Audio 3**, and a full turn-detection stack of wake word, VAD, diarization, and speaker embeddings. NVIDIA's **Audio2Face-3D** even drives avatar facial animation from speech. Two structural choices make this practical rather than merely broad. **Every model is its own SPM library**, so importing only what you need avoids paying for the rest of the catalog in binary size — a real concern for an iOS app. And there is a genuine CLI alongside the Swift API: `brew install speech` gives `speech transcribe`, `speech speak`, `speech translate`, `speech respond`, `speech voice-chat`, and a `speech-server` exposing OpenAI-compatible `/v1/realtime` and `/v1/audio/transcriptions` endpoints locally. Model weights download from HuggingFace on first use and cache locally, with an `offlineMode` flag on every `fromPretrained()` entry point once cached. The README lists **16 public repositories** with verifiable package references, including AnythingLLM. The limits follow directly from the design. This is **Apple Silicon only** — Homebrew installation requires native ARM Homebrew, and Rosetta or x86_64 Homebrew is explicitly unsupported, so there is no Linux, Windows, or Intel Mac story. The Apache-2.0 license covers the library, not the weights: **F5-TTS ships under a non-commercial license and Higgs TTS 3 under a research/non-commercial license**, so anyone shipping a commercial product needs to check each model individually rather than assume the repository license carries through. Version numbers still in the 0.0.x range, with three releases inside two days, signal an API that has not settled.