qwen3-asr-swift ships full local speech stack

This is the kind of project that makes Apple Silicon look like a serious edge-AI speech platform rather than just a good Whisper laptop.

• –The project goes beyond single-model demos by exposing 11 models behind shared Swift protocols, which makes pipeline composition a real engineering feature instead of a README promise
• –Splitting workloads between MLX and CoreML is the sharpest idea here, because it targets the actual bottleneck in local speech apps: resource contention between always-on audio tasks and larger generative models
• –The inclusion of diarization, enhancement, alignment, CLI tools, and an HTTP server makes this feel closer to a deployable speech stack than a narrow model wrapper
• –Benchmarks like sub-real-time ASR and low-latency streaming TTS matter because they make the repo useful for product builders, not just ML hobbyists
• –If the maintainer lands the roadmap items around meeting transcription, streaming diarization, and OpenAI-compatible audio APIs, this could become a strong alternative to fragmented Apple-side speech tooling