具陣列拓樸向量校正之多重訊號分類演算法於即時語音處理多聲源切音與分離

Abstract

本論文提出一套即時的語音資訊處理架構，藉由校正麥克風陣列之拓樸向量(Array Manifold Vector)以後，並偵測聲源方向並追蹤，且實現多聲源切音與分離的方法。本方法結合了多重訊號分類演算法(Multiple Signal Classification)，對聲源頻譜及空間分佈進行估測，並對頻譜中可能為聲源的方向進行機率決策，並利用波束形成原理將不同的方向上語音進行切音與分離。可在對多聲源進行追蹤並保有強健的偵測率，且可排除在聲源頻譜中錯誤偵測的聲源方位。

A real-time system structure for multiple sound sources segmentation and separation using Multiple Signal Classification algorithm is proposed in this thesis. 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. And system structure using the Multiple Signal Classification algorithm to detect and estimate the localization of sound source’s spectrum distribution. And then using probability decision method to determine the direction of sound sources. Finally, multiple sources were extracted from array signals by using beamforming method. This proposed method can track and separate multiple sources at the same time and maintain high detection rate.