Improving multichannel speech enhancement through accurate room-acoustic simulations
The study compared multichannel speech enhancement models trained on different room-acoustic simulations: simplified geometrical acoustics versus high-fidelity wave-based and hybrid simulations. Training with high-fidelity data, which accurately modeled room acoustics, resulted in up to a 38% reduction in word error rate compared to lower-quality training sets. This shows that using physically accurate simulation data directly improves the performance of speech enhancement models without changing the neural network architecture.