Time delay estimation based on pitch harmonics weight and wiener gain for robust speaker localization

Abstract

Speaker localization based on time difference of arrival (TDOA) approach is commonly used in practical with low complexity . In this approach, time delays are estimated between pairs of microphones through signals received by microphones. The estimated time delays can be processed to find the location of speaker. Therefore, it is very essential to obtain good estimated time delays for effective speaker localization. But noise and room reverberation in real environments, affect on exact time delay estimation. Many conventional methods have tried to overcome this adverse environments. Usually the well-known methods for time delay estimation are based on generalized cross correlation (GCC) that is the cross correlation of filtered versions in frequency domain. Especially GCC-PHAT, GCC-ML, corresponding to each filter (called weight function) have different properties each other for noise and reverberation. In this thesis, Analyzing and Compensating these, proposed method is developed. It uses the spectral feature of speech production, called pitch, advantages of GCC methods and Wiener gain for a novel weight function. Through experiments, proposed method is compared with conventional methods and interpretation of proposed method is described.