MSQAT: A multi-dimension non-intrusive speech quality assessment transformer utilizing self-supervised representations

Abstract

Convolutional neural networks (CNNs) have been widely utilized as the main building block for many non-intrusive speech quality assessment (NISQA) methods. A new trend is to add a self-attention mechanism based on CNN to better capture long-term global content. However, it is not clear whether the pure attention-based network is sufficient to obtain good performance in NISQA. To this end, a framework named Multi-dimension non-intrusiveSpeech Quality Assessment Transformer (MSQAT) is proposed. To strengthen the interactions of various speech regions between local and global, we proposed the Audio Spectrogram Transformer Block (ASTB), Transposed Attention Block (TAB) and the Residual Swin Transformer Block (RSTB). These three modules employ attention mechanisms across spatial and channel dimensions, respectively. Additionally, speech quality varies not only in different frames, but also in different frequencies. Thus, a two-branch structure is designed to better evaluate the quality of speech by considering the weighting of each patch's score. Experimental results demonstrate that the proposed MSQAT has state-of-the-art performance on three standard datasets (NISQA Corpus, Tencent Corpus, and PSTN Corpus) and indicate that the pure attention model can achieve or surpass the performance of other CNN-attention hybrid models.