高解析度之國語類音素單元端點自動標示

Sample-based Phone-like Unit Automatic Labeling in Mandarin Speech

在本論文中提出一種以取樣點為單位（sample-based）的高時間解析度之音素端點自動標示與切割的方法，有別於傳統分析語音信號以音框為單位（frame-based）或是音段為單位（segment-based）的研究。本文中，我們提出了一些以取樣點為單位的聲學參數；由實驗結果顯示，這些聲學參數在不同發音特徵之音素轉換間有明顯的變化率，有利於音素切割位置之標記。我們利用這些發音特徵變化的聲學參數特性，建立一個高時間解析度的自動音素端點標示與切割系統。由TCC-300國語語料庫進行自動端點標示之實驗結果顯示，本論文所提出的方法比傳統以音框為單位之切割方法，亦即HMM之切割方法，更能有效切出精準的短停頓、摩擦音、塞擦音等之音素端點位置。

This paper presents a sample-based phone boundary detection algorithm which can improve the accuracy of phone boundary labeling in speech signal. In the conventional phone labeling method adopted the frame-based approach, some acoustic features, like MFCCs, are used. And, the statistical approaches are employed to find the phone boundary based on these frame-based features. The HMM-based forced alignment method is most frequently used method. The main drawback of the frame-based approach lies in incapability of modeling rapid changes in speech signal; moreover, the time resolution of this approach is too coarse for some applications. To overcome this problem, a sample-wise phone boundary detection framework is proposed in this study. First, some sample-wise acoustic features are proposed which can properly model the variation of speech signal. The simple-based spectral KL distance is first employed for boundary candidates pre-selection in order to reduce the complexity of sample-based methods. Then, a supervised neural network is trained for phone boundary detection. Finally, the effectiveness of the proposed framework has been validated on automatic labeling of TCC-300 speech corpus.