FastASR

Why we included this project

FastASR is a C++ inference engine for speech recognition that keeps the runtime footprint deliberately small: just two third-party libraries, FFTW and OpenBLAS, with no PyTorch, Paddle, or TensorFlow involved. That matters most on ARM hardware like the Raspberry Pi 4B, where the full deep-learning frameworks are awkward to install and too slow for real-time use. The project bundles four models, three non-streaming and one streaming, including Paraformer and Conformer variants trained on WenetSpeech and Alibaba's private corpus, so Chinese and mixed Chinese-English recognition quality holds up well. It ships a static library for C++ callers and a pip-installable Python module, and VAD support lets the non-streaming models handle long audio segments. For teams putting voice interfaces on embedded or edge devices, the small footprint and CPU-focused optimization are the main appeal.