CoFi-Lite: Ultra-lightweight speech enhancement using just 83k parameters

CoFi-Lite, developed by researchers Leyan Yang, Dahan Wang, Xiaobin Rong, Jiadong Zhao, and Jing Lu, pushes the boundaries of ultra-lightweight speech enhancement for edge deployment. The model introduces a dual-stream architecture that decouples spectral modeling into coarse-grained (full-band envelopes) and fine-grained (low-frequency details) paths. Two parallel, symmetric encoder-decoder paths extract complementary features, while a Cross-Path Fusion (CPF) module bridges them for efficient feature interaction. This design achieves extreme efficiency: only 83.12k parameters and 12.87M multiply-accumulate operations per second (MACs/s). The paper is accepted by IEEE Signal Processing Letters and includes audio examples.

Performance benchmarks show CoFi-Lite outperforms the established ultra-lightweight baseline GTCRN while consuming just 40.26% of its computational resources. A scaled-up variant of CoFi-Lite delivers performance on par with the current state-of-the-art ultra-lightweight model AdaptCRN, but with a 19.34% reduction in computational cost. These results demonstrate that the coarse-fine decoupling strategy coupled with cross-path fusion is a highly effective approach for pushing the limits of efficiency in speech enhancement. For developers building real-time audio processing on edge devices like smartphones, hearing aids, or IoT microphones, CoFi-Lite offers a compelling path to high-quality noise reduction with minimal power and memory footprint.