A Non-Intrusive Speech Quality Assessment Method Based on Skip-Connection Convolution Architecture and Transformer Blocks

Online conferencing is increasingly popular in everyday communication and office work. However, various noise and transmission distortions directly impact the quality of communication and reduce the quality of experience (QoE). Communication service providers are seeking an accurate and non-intrusive speech quality assessment (SQA) method to promptly identify and address speech issues within communication systems, thereby enhancing their competitiveness in the market. Designing a reliable SQA method faces several challenges: mathematical models inadequately represent human perception, clean reference signals are hard to obtain and the mean opinion score (MOS) rating is time-consuming and labor-intensive. To address these issues, we propose a novel deep learning-based non-intrusive SQA method to predict MOS, termed SAT-SQA. It employs skip-connection convolution architecture as the backbone and transformer blocks as the neck architecture to better capture local and global features. Score prediction is conducted using an attention-based approach, which assigns weights to each speech frame to obtain a weighted global feature, and finally, a fully connected layer is used to produce the final score. We conduct experiments on three publicly available standard datasets from the Conferencing Speech 2022 Challenge: the NISQA corpus, Tencent corpus and PSTN corpus, aiming to predict MOS to closely approximate human scoring. The experimental results indicate that, compared to the current state-of-the-art methods, there is an average reduction of 17.7% in RMSE and an average increase of 4.2% in both PLCC and SRCC.