Sherpa-ONNX STT

From the k2-fsa group behind next-generation Kaldi, sherpa-onnx packages speech models as ONNX files and runs them through ONNX Runtime offline on servers, phones, single-board computers, and in browsers through WebAssembly. For recognition it supports both streaming mode, with text appearing mid-sentence, and decoding of finished recordings, with prebuilt models that include Zipformer transducers, Paraformer, Whisper, NVIDIA NeMo models such as Parakeet, SenseVoice, Moonshine, and others across many languages. Voice activity detection, keyword spotting, punctuation, speaker diarization, and language identification come in the same library, which also does text-to-speech. APIs cover a dozen languages, from C++ and Python to Swift, Kotlin, Dart, and Rust, and a WebSocket server is included. CPU inference is the norm, with optional acceleration on Jetson boards and several NPUs (Rockchip, Qualcomm, Ascend, Axera, Intel). Install via pip install sherpa-onnx or prebuilt binaries. The code is Apache 2.0, while each model keeps its own license. Aimed at developers needing private, on-device transcription.