LIRA: Local Inference tool for Realtime Audio
A CLI-first tool for running and serving automatic speech recognition models locally, with a Ryzen AI NPU backend as well as CPU and GPU, for developers integrating local speech recognition into their own apps.
LIRA is a command-line tool for local, efficient speech recognition. Its lira run command runs, exports or benchmarks a model directly, and lira serve launches an OpenAI-compatible server so applications can call local speech recognition the same way they would call a hosted API.
It supports several ASR model architectures, including different Whisper sizes and the Zipformer streaming encoder, across CPU, GPU and, on AMD Ryzen AI 300 series laptops, NPU acceleration through Vitis AI export artifacts. Installation is a pip-based editable install into a conda environment set up for Ryzen AI.