Show HN: SayItFlow – On-device Mac dictation that unloads models when idle

SayItFlow, a macOS dictation application, launched on Hacker News as a Show HN project. The app runs speech recognition entirely on-device and manages memory by unloading models when the system is idle, allowing users to perform live dictation without cloud dependencies or persistent GPU consumption.

The project was posted by developer innovatorved on the GitHub repository. The core mechanic addresses a common friction point in local speech models: they typically remain loaded in memory or on the GPU even during idle periods, consuming resources unnecessarily. SayItFlow's approach is to load the speech recognition model only when needed and unload it when the system becomes inactive, reducing the memory footprint of an always-available dictation tool.

On-device speech recognition has become more viable for consumer use in recent years as models have gotten smaller and more efficient. Most Mac users default to the built-in dictation, which sends audio to Apple's servers. Third-party on-device solutions eliminate that privacy trade-off but have historically required the user to accept either persistent resource usage or the latency of on-demand model loading. SayItFlow's unload-on-idle strategy attempts to split the difference, keeping the convenience of instant dictation while freeing system resources during gaps in use.

The project remains early, with zero comments on the Hacker News thread as of posting, indicating it is newly launched. No information about future development, monetization, or broader platform support (Windows, Linux) has been announced in available sources.