What inspired me to build this: I use push-to-talk transcription at work every day and wanted the same thing on my personal machine. Every option was paid, $10/month or $30 one-time, and it bugged me because the AI model doing the actual work is free and open source. I mentioned it at dinner with friends, everyone shrugged it off, and that response annoyed me enough to go home and start building. What problem I'm trying to solve: The keyboard has been the bottleneck between human thought and machines for 150 years. We speak roughly 3x faster than we type, but we've been stuck with this interface because we didn't have the technology to turn speech directly into writing. Now we do. Screamer is about making voice the default input method, running entirely on-device, for free, with no cloud dependency or subscriptions. How the approach evolved: Day 1 was just getting it working at all, a Rust prototype with whisper.cpp bindings and a global hotkey that pasted text from my clipboard. Ugly but functional. Day 2 shifted to performance, I added adaptive audio context sizing so the model only processes the actual utterance length instead of the full window, hardware-aware tuning for Apple Silicon vs Intel, and live preview while speaking. That's when latency went from noticeable to instant. Day 3 was the unglamorous but necessary packaging work, Apple code signing and notarization to make it a proper .app. Days 4 and 5 were landing page and launch. The core technical insight that emerged during building was the dual pipeline architecture, keeping two separate Whisper states warm so live preview and final transcription never block each other.

Screamer

The fastest free speech to text AI in the world.

The fastest free speech to text AI in the world.