Speech Based Onset Estimation for Multisensor Localization

Abstract

This work presents a study of a current problem in the field of audio processing: Source and receiver localization. Currently, this problem requires that either the onset time of the sources or the internal delay of the receivers are known. The algorithms studied here, take advantage of the structure of the time matrix, which contains the TOA of all the receivers with respect to all the sources, and finds the solution to the locations when the onset times are known. The problem here is then approached from a time difference of arrival (TDOA) perspective, which inherently cancels the onset times by subtracting the time of arrival (TOA) of a source at every receiver. An alternative approach is also proposed, which uses speech signals as calibration signals in order to estimate the onset times. Such an approach is based on an algorithm which uses artificial calibration signals to calculate the onsets. Those signals are known a-priori, which implies that an additional device which produces those signals is needed. Once both internal delays and onset times are known, the locations of both sources and receivers can be estimated using a current algorithm which is also described here