Hybrid solutions for the one-microphone speech enhancement problem

Incorporating (hand-crafted) structured models into traditional speech enhancement approaches gives an improvement, albeit constrained by the model's limitations. Alternatively, employing deep neural networks (DNNs) leads to better performance for conditions seen during training. However,
these methods generalise poorly. We propose to systematically incorporate structured knowledge into DNNs, thereby combining significantly improved speech enhancement with greater robustness to unseen conditions.