基于张量深度学习和注意力机制的无参考声音质量评价方法研究

无参考声音质量评价提供了一种非插入式的衡量用户听音体验的客观化技术手段，是语音传输和音频回放系统性能的重要评估指标。未来以IP为基础的高清语音以及沉浸式的空间音频成为语音和音频系统的主流，如何有效评价无法获取参考音源情况下的输出声音质量成为行业急需但不容易解决的问题。本研究将以准确反映用户主观听觉体验为最终目标，致力于从人因维度出发对大规模声音数据的损伤特征进行高阶空间上的深度感知和融合，利用张量深度学习和注意力机制对受损声音信号的听觉体验机理进行建模，建立从感知到认知的两级学习模型并进行联合优化，进一步引入迁移学习将无参考语音质量评价扩展到空间音频质量评价。本项目从提高语音和音频系统用户听觉体验满意度的实际需求出发，引入新的技术手段解决无参考声音质量客观评价这一具有挑战性的问题，为语音和音频评价技术领域提供新的建模方法和研究思路，具有重要的理论和应用意义。

The non-reference sound quality evaluation provides a non-intrusive objective technical measure of the user's listening experience, and is an important evaluation index for the performance of speech transmission and audio playback systems. In the future, IP-based high-definition voice and immersive spatial audio will become the mainstream of voice and audio systems. How to effectively evaluate the output sound quality in the case where a reference sound source cannot be obtained has become a problem that is urgently needed but not easily solved in the industry. In this research, the ultimate goal is to accurately reflect the user's subjective auditory experience. It is committed to the depth perception and fusion of the damage characteristics of large-scale sound data in high-order space from the human factor dimension. The auditory experience mechanism of the impaired sound signal is modeled by use of tensor deep learning and attention mechanism, a two-level learning model from perception to cognition is established for joint optimization, and transfer learning is further introduced to extend the evaluation of non-reference speech quality to the evaluation of spatial audio quality. This project introduces new technical methods to solve the challenging problem of non-reference sound quality objective evaluation from the actual needs of Improving user hearing experience satisfaction with speech and audio systems. Our research will provide new modeling methods and research ideas for the technical field of speech and audio evaluation, which has important theoretical and application significance.