Hybrid real-complex network improves speech enhancement with fewer ops
Complex-valued speech models get a lean hybrid makeover for cleaner audio...
Complex-valued neural networks are powerful for processing time-frequency representations in speech enhancement, but they often carry a heavy computational burden — especially in small models. To address this, researchers led by Luan Vinícius Fiorio propose a hybrid architecture that pairs a real-valued magnitude-mask branch with a complex-valued additive correction branch. The two branches are coupled via domain conversion functions at the bottleneck, enabling the model to get the benefits of complex-domain processing without the full computational overhead.
Their approach was applied to two network designs: a convolutional denoising autoencoder (CDA) and a convolutional-recurrent network (CRN). Averaged over four signal-to-noise ratios (SNRs), the hybrid models improved both intelligibility and perceptual quality compared to real-only, complex-only, and matched-parameter baselines — while substantially cutting the number of operations. The results suggest that hybrid real-complex architectures are an efficient middle ground for speech enhancement, making them practical for real-time and edge-device deployments.
- Hybrid architecture: real magnitude-mask branch + complex additive correction branch, coupled at the bottleneck
- Validated on convolutional denoising autoencoders and convolutional-recurrent networks across 4 SNRs
- Improved intelligibility and quality while substantially reducing computational operations vs. complex-only models
Why It Matters
Efficient hybrid speech enhancement makes high-quality audio processing viable for real-time edge devices like hearing aids and smart speakers.