Short Lithuanian commands can use a recognizer that processes the audio received instead of Whisper’s fixed-length window. Wyoming Faster Whisper selects the LT_ausys model, based on Datolabs’ Upelis, when its optional onnx-asr speech engine is installed.
In the contributor’s 127-command test on a Ryzen computer, median recognition time was 0.18 seconds, compared with 3.55 seconds for the tested Paprika Whisper model. The recordings came from one speaker. This changes the standalone server package; it does not establish availability through the Home Assistant OS Whisper app.