Multilevel B-Splines-Based Learning Approach for Sound Source Localization

In this paper, a new learning approach for sound source localization is presented using ad hoc either synchronous or asynchronous distributed microphone networks based on the time differences of arrival (TDOA) estimation. It is first to propose a new concept in which the coordinates of a sound source location are defined as the functions of TDOAs, computing for each pair of microphone signals in the network. Then, given a set of pre-recorded sound measurements and their corresponding source locations, the multilevel B-splines-based learning model is proposed to be trained by the input of the known TDOAs and the output of the known coordinates of the sound source locations. For a new acoustic source, if its sound signals are recorded, the correspondingly computed TDOAs can be fed into the learned model to predict the location of the new source. Superiorities of the proposed method are to incorporate the acoustic characteristics of a targeted environment and even remaining uncertainty of TDOA estimations into the learning model before conducting its prediction and to be applicable for both synchronous or asynchronous distributed microphone sensor networks. The effectiveness of the proposed algorithm in terms of localization accuracy and computational cost in comparisons with the state-of-the-art methods was extensively validated on both synthetic simulation experiments as well as in three real-life environments.