k2-fsa/sherpa-onnx

Built on ONNX Runtime with Kaldi-style FST-based decoding, the project enables both streaming and non-streaming inference across diverse architectures including specialized NPU accelerators (Rockchip, Qualcomm, Ascend, Axera) for hardware-optimized performance. Beyond core speech tasks, it includes specialized modules like keyword spotting, audio tagging, spoken language identification, and punctuation restoration—accessible through a unified C++/ONNX backend with language bindings spanning Python, Go, Java, Rust, and WebAssembly. Pre-trained models are available on Hugging Face with browser-based demos, supporting multi-language inference (Chinese, English, Cantonese, Japanese, Korean, Thai) on everything from Raspberry Pi to NVIDIA Jetson edge devices.

10,885 stars and 181,098 monthly downloads. Actively maintained with 138 commits in the last 30 days. Available on PyPI.