unramble

so by the time you release the key the transcript is mostly done. Two independent WebSocket connections race each other, and if both fail, an HTTP batch fallback catches it. The transcript goes through a post-processing step that removes filler words and fixes grammar. About 40% of dictations are clean enough to skip this step entirely. When post-processing is needed, a fast model handles it in about 0.4 seconds. It's designed to be taken apart and reassembled. You can swap the speech model, rewrite the prompts, add new languages, or fork the entire experience to fit how your team works. I'm hoping people will morph it into other products. The FreeFlow service is open source. You can self-host it, but running a low-latency streaming dictation service for a team is real infrastructure work: persistent WebSocket connections, streaming routes to speech models, failover, rate limits. At a company with fifty or five hundred people, keeping that reliable is a job in itself. FreeFlow uses Autonomy to make this easy. On first launch, the macOS app deploys the service to a private server. Two minutes, no infrastructure knowledge needed. You can then invite your team. One server handles everyone, no per-seat fees. It sustains thousands of simultaneous streaming connections. In a stress test, 50 people dictating at the same time got sub-second latency with zero failures. brew install build-trust/freeflow/freeflow It's macOS only for now, but I plan to build for other operating systems. The two most useful contributions right now are mic compatibility data (every mic behaves differently) and prompts that improve polish quality for a specific language. Try it, tell me how it works with your mic and your apps. What's fast, what's slow, what's broken. GitHub: https://github.com/build-trust/freeflow