A Deep CNN Self-Attention Model for Multidimensional Speech Quality Prediction Using Crowdsourced Datasets

Predicting voice quality is critical for improving communication systems and the user experience. In this article, we offer a unique technique to multidimensional speech quality prediction using a Deep Convolutional Neural Network (CNN) enriched with self-attention techniques and crowdsourced datasets. The proposed model incorporates both spatial and temporal information from speech signals via numerous convolutional layers, enabling it to accurately represent the intricate patterns and features seen in speech data. Furthermore, self-attention methods are used to let the algorithm to focus on relevant parts of the input data, increasing its discriminative capacity and tolerance to interference. Transfer learning is utilized to leverage pre-trained representations from large-scale speech datasets, enabling the model to generalize well to diverse speech characteristics and environments. By fine-tuning the pre-trained model on the specific crowdsourced dataset, we adapt it to the task of speech quality prediction while mitigating the need for large annotated datasets, which are often costly and time-consuming to obtain. Experimental results demonstrate the effectiveness of the proposed approach, achieving state-of-the-art performance on various metrics for multidimensional speech quality prediction tasks. Furthermore, extensive ablation studies are conducted to analyze the contributions of different components of the model, highlighting the importance of both the CNN architecture and self-attention mechanisms. Overall, our findings suggest that the proposed deep CNN self-attention model, combined with transfer learning from crowdsourced datasets, presents a promising approach for accurate and efficient speech quality prediction in diverse real-world scenarios