FastEnhancer: Speed-Optimized Streaming Neural Speech Enhancement (2509.21867v1)

Abstract: Streaming speech enhancement is a crucial task for real-time applications such as online meetings, smart home appliances, and hearing aids. Deep neural network-based approaches achieve exceptional performance while demanding substantial computational resources. Although recent neural speech enhancement models have succeeded in reducing the number of parameters and multiply-accumulate operations, their sophisticated architectures often introduce significant processing latency on common hardware. In this work, we propose FastEnhancer, a streaming neural speech enhancement model designed explicitly to minimize real-world latency. It features a simple encoder-decoder structure with efficient RNNFormer blocks. Evaluations on various objective metrics show that FastEnhancer achieves state-of-the-art speech quality and intelligibility while simultaneously demonstrating the fastest processing speed on a single CPU thread. Code and pre-trained weights are publicly available (https://github.com/aask1357/fastenhancer).