Research On Speech Source Localization Methods Based On Microphone Arrays

Microphone arrays (MA) are widely used in audio/video conferences, speech recognitions, and speech enhancements etc. The localization of the speech source is the primary task of the array signal processing, and the basis of spatial filter designing. The source localization strategies include the high (super)-resolution algorithms and the general resolution ones. To the general resolution method, the precision of localization is limited by the array aperture, and has only been used in low precision case. Used the high resolution algorithm, the precision of localization can get beyond the Rayleigh Resolution Limit decided by the array aperture, and can even gain arbitrary resolution at some case. Therefore, this type of methods is of great value. The classical high-resolution methods suppose that the signal sources are narrow-band and stationary in the far field. However, the speech signal is wide-band and short-time stationary in the near field. And then, the classical DOA (direction of arrival) estimation methods can't solve the speech source localization problems. Speech source multi-dimension localization (MDL) methods based MA, vs. classical DOA methods, have some problems as follows:1) Wide-band signal: In the case of narrow-band signal, the phase-difference between two adjacent elements of array is supposed to be a function of source's location, and the frequency of the signal is a constant. While, speech signal is wide-band and non-modulated, whose phase-difference is a compound function of the frequency and the location of sources.2) Near field source: In the MA processing, speech source is usually the near field of the array. While, the source lies in the far field of the array in the classical array processing.3) Spatial interference signal: The spatial interference and speech signal are captured simultaneously by the MA in the room environment, resulting in the bad performance of the speech source localization4) Multi-dimension localization: it requires two- or three-dimension localization in MA application, but only one-dimension localization in the classical array processing.Focusing on these problems, several algorithms of speech source localization are presented as follows:1) The near field signal model based on MA: To meet the array signal...