Show HN: I built a local macOS dictation app using Nvidia Parakeet and MLX(sayline.app)
sayline.app
Show HN: I built a local macOS dictation app using Nvidia Parakeet and MLX
https://www.sayline.app/
2 comments
I am super bullish on bringing these experiences to Mac users. Keep it up.
Tips: Ditch the purple. Sarah Chen is Fakey Fakerson.
Tips: Ditch the purple. Sarah Chen is Fakey Fakerson.
Thanks for the reply! Fakey Fakerson got me so bad lol. Which color would you replace the purple with?
The app uses NVIDIA's Parakeet v3 (TDT) as the primary engine. Inference is handled by the FluidAudio library. It's insanely fast on M-series chips. For Intel users, I added local Whisper support ranging from Tiny up to Large v3 Turbo (quantized to save space).
The interesting part was the AI integration. I originally tried to pipeline the ASR output into Apple's Intelligence Foundation Model. It was a nightmare, because the content moderation filter blocked nearly every useful request.
So I scrapped that and moved to SwiftMLX. Now it runs Qwen 3.0 (0.6B to 8B) locally. You just set a hotkey, speak, and the local LLM handles formatting/rewriting/translation with very low latency. Zero data leaves your Mac.
It's a one-time purchase (no monthly fee). I'd love to hear what you think about the latency compared to standard Whisper implementations.