Research On Individualized Modeling Method Of Head Related Transfer Function

The head related transfer function(HRTF)describes the filtering process of sound waves from the sound source to the binaural,and the HRTF is used to filter the audio signal to reproduce the realistic binaural three-dimensional audio effect.The HRTF is closely related to the human body's morphological features(head,torso,auricle,etc.).The use of non-personalized HRTF by the listener can lead to confusion,confusion,and inaccurate positioning.Therefore,obtaining a personalized HRTF suitable for listeners has become an increasingly important issue in the research and practice of binaural three-dimensional audio.Aiming at the problem that the existing HRTF personalization technology establishes the mapping relationship between human morphological features and HRTF through linear regression method,it can not express the complex nonlinear relationship between them.This paper uses neural network to establish the relationship between them and achieve personalized HRTF generation.Aiming at the problem that the reduction of HRTF by principal component analysis(PCA)will lead to the loss of high-frequency detail information with small energy components in HRTF,a common pole/zero(CPZ)model was proposed to parameterize HRTF,and establish a HRTF personalized generation method.The main work and results are as follows:(1)The PCA was used to reduce the dimension of the amplitude spectrum of the HRTF,reducing redundancy in the original data while preserving as much as possible of the changes in the data.Correlation analysis was used to screen human morphological features to remove highly correlated features.A radial relationship(RBF)neural network was used to establish a mapping model between human morphological features and HRTF.Objective experiments show that the average spectral distortion of the personalized HRTF generated based on the RBF neural network is reduced by 0.53 dB compared with the linear regression method.Subjective experiments show that the personalized HRTF predicted by the personalized modeling method proposed in this paper is better than the general HRTF,and the average CMOS score is 1 point higher than the general HRTF.(2)The HRTF is parametrically expressed using the common pole/zero model,and the key auditory clue information(peaks and troughs)in the HRTF is retained.The RBF neural network is used to establish the mapping model between human morphological features and parameters in the CPZ model.Objective experiments show that The spectral distortion of the personalized HRTF generated based on the CPZ and RBF neural networks is 0.74 dB lower than that of the linear regression method.Subjective experiments show that the personalized HRTF predicted by the personalized modeling method proposed in this paper is better than the general HRTF,and the average CMOS score is 1.13 points higher than the general HRTF.The generation method of the personalized head related transfer function proposed in this paper can realize the low-cost and high-efficiency generation of personalized head-related transfer functions through human morphology features,and has a good application prospect in practical applications.