Target Speech Signal Extraction Using Variance of Phase Difference

It has been attracted attention to a hearable device which is one of wearable devices on human ears for communications, entertainment, health coaching, and so on. For human-to-human speech communication, an important issue of the hearable device is to enhance a target speech signal in an environment where multiple speech signals exist. In this paper, we propose a method to extract the target speech signal under the assumption that the target speaker be the nearest and the front to the hearable device user. The proposed method separates an observed speech mixture to each speech signals and estimates the target speech signal based on a variance of a phase difference between both ears. The proposed method can be implemented as a batch algorithm or an online algorithm. We evaluate the capability of the proposed method on both algorithms, and show that the online algorithm is more effective to a hearable device.