Leveraging the NOMAD and Intelligibility Loss to Improve MP-SENet for Speech Enhancement

The aim of this study is to enhance MP-SENet, a highly effective speech enhancement network that denoises both magnitude and phase spectra concurrently. The MP-SENet model is trained utilizing multi-level losses on magnitude spectra, phase spectra, short-time complex spectra, and time-domain waveforms. In this work, we suggest that further refinement of the loss function could be achieved through the integration of STOI loss and NOMAD loss. STOI assesses the objective intelligibility of the enhanced signal, while NOMAD is an audio metric that can be applied to any non-matching reference and is perceptually differentiable. The results obtained from the preliminary experiments performed on the VoiceBank-DEMAND task indicate that the incorporation of STOI and NOMAD loss into the MP-SENet training procedure leads to a substantial improvement in various SE metric scores for the test dataset.