This repository implements ASR inference in C++ with minimal dependencies, optimized for fast execution on resource-constrained platforms including ARM devices like Raspberry Pi 4B. It supports four models: paraformer (Paraformer-large), k2_rnnt2 (pruned transducer stateless2), conformer, and conformer_online. Models are derived from Google’s Transformer architecture and trained on WenetSpeech (10000+ hours) and private datasets (60000+ hours), offering both streaming and non-streaming recognition modes.