ced.cpp Brings Xiaomi's Audio-Tagging Models to C++17 via ggml, No Python Required

The LocalAI team has released ced.cpp, a C++17 inference port of Xiaomi's CED (Consistent Ensemble Distillation) audio-tagging models built on the ggml runtime. The library classifies audio into any of the 527 AudioSet categories — from baby cries to glass breaking — and runs entirely on CPU with no Python dependency at inference time. Model variants range from the 5.5M-parameter ced-tiny (a 6 MB GGUF file capable of running on Raspberry Pi-class hardware) up to the 86M-parameter ced-base, all distributed through the mudler/ced-gguf collection.

Benchmarked against PyTorch on an AMD Ryzen 9 9950X3D, ced.cpp in f32 mode processes a 10-second clip in 126.6 ms versus PyTorch's 158.8 ms, while using roughly half the peak RAM. Quantized q8_0 inference drops memory further to 111 MB. Numerical parity with the upstream PyTorch models is validated end-to-end to within ~2.4e-7 in f32, preserving identical top-5 tags. The project also wires directly into LocalAI as a REST and real-time WebSocket backend for live sound-event recognition.