faster-enhancer.c
A dependency-free C and SIMD int8 runtime that runs the FastEnhancer-Medium speech enhancement model in real time, without an inference framework or heap allocation after startup.
faster-enhancer.c is a C/SIMD runtime for FastEnhancer-Medium, a streaming speech enhancement model, running at 48 kHz. It runs the published FastEnhancer weights unchanged and is training-free: it uses post-training int8 quantization with activation ranges recomputed from each frame rather than a calibration set.
The runtime is streaming and causal, processing 320 samples in and 320 out with no look-ahead beyond a fixed STFT alignment delay. All state lives in one fixed-size structure so long runs do not trigger further allocation, and six hand-written int8 GEMM kernel tiers are selected at initialization based on the host CPU's instruction set, with no scalar fallback. Output is byte-identical across kernel tiers within an architecture family.
The public API is four C functions: fe_init, fe_run, fe_reset and fe_free. It builds with CMake into a static library, plus command-line runners for trying it on real audio.