chenkui164/FastASR

Implements four distinct model architectures—three non-streaming (Paraformer, K2-RNNT2, Conformer) and one streaming (Conformer Online)—with VAD-based long-form audio support, using memory-mapped binary model files for efficient parameter loading. Built entirely in C++ without deep learning framework dependencies, relying only on libfftw3 and libopenblas, with pointer-based algorithms that minimize reshape/permute operations to maximize CPU utilization across x86, ARM, and other platforms. Provides both a static C++ library and Python bindings across CPython 3.6-3.11, supporting Linux, macOS, and Windows with pre-built wheels.

547 stars. No commits in the last 6 months.