Auditory Scene-Attention Model For Speech Enhancement

In this work, we propose a new speech enhancement model referred to as auditory scene-attention model (ASAM), that can adapt dynamically to changes in the auditory scene components, such as speaker gender, input SNR levels, and background noise properties. To this end, a representative set of so-called Universal Scene Models (USM), each associated to a different auditory scene component, are first created, where each model attempts to predict a corresponding ideal ratio mask (IRM). The dynamic adaptation to changes in the auditory scene is then carried by computing the outputs of the USMs and forming a weighted combination of the most relevant scene models. This adaptation process is implemented via a frame-based attention mechanism, allowing to realize a soft selection of USM models, and taking advantage from both scene-dependent and scene-independent models. The evaluation of the proposed ASAM model, under different noise conditions and input SNR levels, shows substantial improvements in terms of standard speech Quality and intelligibility measures.