Deep Learning-Based Non-Intrusive Multi-Objective Speech Assessment Model With Cross-Domain Features

This study proposes a cross-domain multi-objective speech assessment model, called MOSA-Net, which can simultaneously estimate the quality, intelligibility, and distortion scores of an input signal. MOSA-Net comprises convolutional neural network bidirectional long short-term memory architecture for representation extraction, multiplicative attention layer fully connected each metric prediction. Additionally, features (spectral time-domain features) latent representations from self-supervised learned (SSL) models are used as inputs to combine rich acoustic information obtain more accurate assessments. Experimental results show that in both seen unseen noise environments, improve linear correlation coefficient (LCC) perceptual evaluation quality (PESQ) prediction, compared Quality-Net, existing single-task model PESQ LCC short-time objective intelligibility (STOI) STOI-Net, STOI Moreover, be pre-trained effectively adapted predicting subjective with limited amount training data. mean opinion score (MOS) predictions, MOS-SSL, strong MOS We further adopt guide enhancement (SE) process derive quality-intelligibility (QI)-aware SE (QIA-SE) approach. QIA-SE outperforms baseline system improved environments over model.