No-reference Speech Quality Assessment of SWB Signal Based on Machine Learning
Ran Wang, Yitong Liu, Guanglei Ding, Yangguang Wei, Hongwen Yang, Wireless Theories, Technologies Lab · 2018
A novel no-reference method based on machine learning algorithm for SWB (Super Wideband) speech quality measurement is proposed in this paper. For this, a SWB audio database including 28786 degraded signal samples acquired from actual network is built. With this database, a multidimensional waveform feature description is built up for the degraded signal. Gradient Boosting Decision Tree (GBDT) algorithm is used to design the model, establishing the relationship between the feature description and the perceptual mean opinion score (MOS). The results predicted by the proposed method coincide with the perceptual quality and demonstrate more effective performances than the current P.563 algorithms, in terms of Pearson coefficient, root mean square error and mean absolute percent error.