HRTF low-dimensional representation based on deep convolutional autoencoder and attention mechanism

Abstract

Head-Related Transfer Function (HRTF) depicts the reflection and scattering effects of the environment and the human body on sound during the transmission of sound signals from the sound source to the human ear and contains a large amount of auditory cue information for auditory localization. Due to the high-dimensional complexity and nonlinear nature of the sample data of HRTF itself, it creates difficulties in analyzing the relationship between the auditory localization cues of HRTF and the spatial orientation and morphological features of the human body. The traditional lowdimensional representation makes it difficult to effectively deal with the complex nonlinear relationships between multiple auditory cues in HRTF, resulting in performance degradation. To solve this problem, this study proposes a low-dimensional representation method for HRTF based on a deep convolutional autoencoder. The method considers that HRTF spectral features have the property of continuous variation in three-dimensional space and integrates the nonlinear relationships of full-space HRTF features by modeling the natural spatial attributes of the HRTF ensemble data. Firstly, the attention mechanism is introduced in the encoder, which solves the bias caused by mapping HRTF to a 3D tensor for convolution operation and mines the intrinsic features implied between the spatial orientations of HRTF neighborhoods and neighboring spectra, which improves the low-dimensional representation ability of the network. Secondly, the combination of dense connectivity and attention mechanism in the decoder according to the characteristics of different levels guarantees the effective delivery of lowdimensional features. Experimental results on several publicly available HRTF datasets show that the proposed model outperforms traditional methods in the low-dimensional representation and reconstruction of HRTFs and realizes high-performance low-dimensional representation and reconstruction of HRTFs.