[Speech Toolkit] Add a MeetingTranscriber pipeline

Open-source Swift package running 11 speech models on Apple Silicon via MLX (GPU) and CoreML (Neural Engine). Fully local inference, no cloud dependency. Models implemented: **ASR** \- Qwen3-ASR 0.6B/1.7B (4-bit), Parakeet TDT (CoreML INT4) - RTF \~0.06 on M2 Max **TTS** \- Qwen3-TTS 0.6B (4-bit), CosyVoice3 0.5B (4-bit) - Streaming, \~120ms first chunk **Speech-to-speech** \- PersonaPlex 7B (4-bit) - Full-duplex, RTF \~0.87 **VAD** \- Silero v5, Pyannote segmentation-3.0 - Streaming + overlap detection **Diarization** \- Pyannote + WeSpeaker + spectral clustering - Auto speaker count via GMM-BIC **Enhancement** \- DeepFilterNet3 (CoreML) - Real-time 48kHz noise suppression **Alignment** \- Qwen3-ForcedAligner - Non-autoregressive, RTF \~0.018 Key design choice: MLX for large models on GPU, CoreML for small models on Neural Engine. This lets you run VAD on ANE while ASR runs on GPU without contention — something WhisperKit struggles with (their Core ML audio encoder blocks the ANE for 300-600ms per call). All models conform to shared protocols, so you can swap implementations or compose pipelines. Currently working on a MeetingTranscriber pipeline (diarize → per-segment ASR) and streaming real-time diarization. Roadmap: [https://github.com/soniqo/speech-swift/discussions/81](https://github.com/soniqo/speech-swift/discussions/81) Repo: [https://github.com/soniqo/speech-swift](https://github.com/soniqo/speech-swift)