基于层级集成的个性化空间音频技术

Hierarchical integration-based individualized spatial audio technique

个性化的头相关传输函数(head-related transfer function,HRTF)可以有效改善空间音频质量。针对个性化HRTF难以精确获得的问题,提出了一种基于层级集成的个性化空间音频生成方法。该方法通过三个模型逐层建立个性化HRTF中的定位信息。首先,采用高斯混合模型建立用户无关的共用模型。然后,采用自编码器获得与用户有关的HRTF的隐表示,利用深度神经网络在人体生理参数与HRTF的隐表示之间建立非线性映射,得到用户有关的个性化模型。为了尽可能恢复个性化HRTF细节信息,对上述模型降维过程中的残差进行线性建模,得到残差模型。对于目标用户,任意空间位置处的个性化的HRTF可以通过集成三个层次下的模型获得,用于生成三维空间音频。最终,实验结果表明,提出的算法可以有效降低HRTF频谱损失,提升对个性化HRTF的预测性能。

The individualized head-related transfer function(HRTF)can effectively improve the quality of spatial audio.Aiming at the problem of the difficulty of acquiring individualized HRTFs accurately,this paper proposed an individualized spatial audio generation method based on hierarchical integration.The method utilized three models to obtain the location information of individualized HRTFs layer by layer.Firstly,the method built a user-independent common model using the Gaussian mixture model.Secondly,it trained the AutoEncoder to generate the latent representation of user-dependent HRTFs.Then,it used a deep neural network to build the nonlinear mapping between human anthropometric parameters and the latent representation of HRTFs,resulting in a user-dependent individualized HRTF model.In order to recover the spectrum details of the individualized HRTFs as much as possible,it modeled the residual in the process of dimension reduction linearly to obtain the residual model.For target users,the individualized HRTFs at any spatial location could be obtained by integrating the three models and then used to generate the spatial audio.Finally,the experimental results show that the proposed method can reduce the spectrum loss of HRTFs effectively,and improve the predictive performance of individualized HRTF.