具陣列拓樸向量校正之多重訊號分類演算法於多聲源切音與分離

Abstract

本論文提出一套利用校正過之陣列拓樸向量(Array Manifold Vector)，提升多重訊號分類演算法(Multiple Signal Classification)效果在寬頻估測時的準確度，並實現多聲源切音與分離的方法。本方法結合了聲源頻譜與空間分佈資訊，利用機率決策對未知數量聲源方位進行分類，並將不同聲源語音利用波束形成原理進行切音與分離。本方法由於進行了陣列拓樸向量的完善校正，保證在低訊噪比下對多聲源的方位保有相當程度的正確率，且可排除錯誤偵測的聲源方位。

This thesis proposes a system structure for multiple sound sources segmentation and separation using MUSIC (Multiple Signal Classification) algorithm. Using a calibrated array manifold vector, the proposed calibration method improves the accuracy of the MUSIC algorithm for wide-band detections, hence providing high accuracy source segmentation and separation results. The system structure uses a multiple signal classification algorithm to detect the location of sound sources and estimate their spectrum distributions. The multiple sources tracking method is implemented by a probability decision method regarding spatial and spectrum distributions. Using the estimated directivity, multiple sources were extracted from the array signals using beamforming methods. This proposed system structure can track and separate multiple sources at the same time and maintain high detection rate under very low SNR conditions.