I built PulseScribe because I wanted Wispr Flow without the subscription.
Key features:
- ~300ms latency with Deepgram WebSocket streaming
- Local mode with MLX/Lightning on Apple Silicon (M1/M2/M3)
- Context-aware LLM refinement (detects email, chat, code)
- Just fixed it for macOS 26 (Tahoe) this week
Technical stack: Python, PyObjC for native macOS UI, MLX for local inference.
The tricky part was getting <500ms end-to-end latency. We stream audio over WebSocket while recording, so results appear immediately when you stop.
I built PulseScribe because I wanted Wispr Flow without the subscription.
Key features: - ~300ms latency with Deepgram WebSocket streaming - Local mode with MLX/Lightning on Apple Silicon (M1/M2/M3) - Context-aware LLM refinement (detects email, chat, code) - Just fixed it for macOS 26 (Tahoe) this week
Technical stack: Python, PyObjC for native macOS UI, MLX for local inference.
The tricky part was getting <500ms end-to-end latency. We stream audio over WebSocket while recording, so results appear immediately when you stop.
GitHub: https://github.com/KLIEBHAN/pulsescribe Website: https://pulsescribe.me
Impressive latency numbers! I’ll definitely try replacing super whisper with this.