Local, offline speech-to-text service exposing the Myna session API on a
Unix domain socket. Wraps NVIDIA's Nemotron cache-aware streaming models
(English, and 3.5 multilingual) as int8 ONNX served via onnxruntime: every
word is committed about half a second after it is spoken and never revised,
which suits live captions.
The service receives PCM audio pushed by clients; it has no access to the
microphone or any audio hardware.