Hybrid Real- and Complex-Valued Neural Network Architecture for Speech Enhancement

This paper investigates hybrid real- and complex-valued neural networks for monaural speech enhancement. While complex-valued models can process time-frequency representations natively, they often increase computational cost and can be inefficient in small-model regimes. We therefore study a matched-parameter hybrid architecture that combines a real-valued magnitude-mask branch with a complex-valued additive correction branch, coupled via domain conversion functions at the bottleneck. The approach is applied to convolutional denoising autoencoder and convolutional-recurrent network architectures. Averaged over four SNRs, the hybrid models improve intelligibility and quality over the considered counterparts, while substantially reducing the number of operations.