On the use of bottleneck features of CNN auto-encoder for personalized HRTFs

Abstract

The most effective way of providing immersive sound effects is to use head-related transfer functions (HRTFs). HRTFs are defined by the path from a given sound source to the listener’s ears. However, sound propagation by HRTFs differs slightly between people because the head, body, and ears differ for each person. Recently, a method for estimating HRTFs using a neural network has been developed, where anthropometric pinna measurements and head-related impulse responses (HRIRs) are used as the input and output layer of the neural network. However, it is inefficient to accurately measure such anthropometric data. This paper proposes a feature extraction method for the ear image instead of measuring anthropometric pinna measurements directly. The proposed method utilizes the bottleneck features of a convolutional neural network (CNN) auto-encoder from the edge detected ear image. The proposed feature extraction method using the CNN-based auto-encoder will be incorporated into the HRTF estimation approach.