對於公共電視新聞語料之語者與環境轉換偵測

Abstract

在本論文中，我們提出非監督式的語者及環境轉換偵測，其中使用GMM來描述相異兩個聲音片段之統計特性，利用共用的mixture component(Common Component GMM，CCGMM)來減少估計混合權重時之計算量，以估計出之權重向量來代表聲音片段之特性，進而量測相鄰聲音片段間的相異度，決定出可能之轉換點。我們以公共電視新聞語料(PTSND)來測試我們提出之相異度量測方法，並且與傳統的貝氏資訊法則(BIC)之相異度量測方法比較，有著較佳的轉換點之辨識率，其MDR與FAR分別為18.8 %及16.8 %。

In this thesis, a GMM with common mixture components, referred to as the common component GMM(CCGMM), is proposed to be the diversity measure for the speaker and environment change detection in Public Television Service Database (PTSND). The use of GMM is to increase the accuracy of audio signal modeling while the use of common mixture components is to solve the complexity problem of parameter estimation and divergence measure evaluation. Experimental results on the PTSND showed that it outperformed a conventional BIC-based method. A MDR of 18.8% with 16.8% FAR is achieved.