臺灣博碩士論文加值系統

本論文對三種最主要的韻律聲學參數做微結構的初步研究，包括基頻軌跡、音長、能量軌跡，本研究以音素做為分析單位，分析一個包含四種不同語速之平行語料庫的三種音素聲學韻律參數的變化。首先，將此語料庫之語句信號以強迫對準切割成音素序列，並且對此語料庫使用非監督式韻律標記與模式方法，做自動停頓及韻律狀態的韻律標記。研究主要是探討語言參數及兩類韻律標記對音素長度、音素能量軌跡、以及音素的基頻存在機率的影響，方法是使用決策樹，將各種影響因素以問題形式加入，來決定在各種前後文的情況下三種音素聲學韻律參數的估計值，實驗結果顯示此做法可以改善此三種韻律參數的合成。

A preliminary study on the microstructure of Mandarin speech prosody is conducted in this thesis. The variations of three phoneme-based prosodic-acoustic features on a large speech corpus containing four parallel sub-corpora are explored. They include phone duration, phone energy contour, and phone’s occurrence probability of pitch. First, all utterances of the corpus are segmented into phoneme sequences by forced alignment. Meanwhile, labelling of utterance with two prosodic tags of break and prosodic state is performed by the PLM algorithm proposed previously. Then, the influences of linguistic features and prosodic tags on the three prosodic-acoustic features are analyzed. The method of regression tree is employed to determine the estimates of these three prosodic-acoustic features in different contexts considering various combinations of linguistic features and prosodic tags. Experimental results show that the proposed method can improve the syntheses of these three prosodic-acoustic features.

中文摘要……………………………………………………………………………..Ⅰ
Abstract…………………………………………………………………………...….Ⅱ
致謝……………………………………………………………………………….….Ⅲ
目錄……………………………………………………………………………….….Ⅳ
表目錄……………………………………………………………………….……….Ⅶ
圖目錄………………………………………………………………………............ Ⅷ
第一章 緒論 1
1.1 研究動機 1
1.2 研究方向 1
1.3 語料庫簡介 2
1.4 章節概要說明 2
第二章 階層式韻律模型 3
2.1漢語語音階層式韻律架構 3
2.2 韻律模型設計 5
2.2.1音節韻律模型 7
2.2.2停頓聲學模型 8
2.2.3韻律狀態模型 10
2.2.4停頓語法模型 10
第三章 音素聲學參數模型訓練 11
3.1 音素邊界切割 11
3.2 音素長度之語速正規化 12
3.2.1由正規化後的音節長度使用比例分配音素長度 13
3.2.2 phone-based的音素長度分配 13
3.2.3綜合phone-based及syllable-based音節長度正規化方法 15
3.3音素長度之估計 16
3.3.1 HTS長度分配公式的做法 16
3.3.2未經語速正規化之音素長度預估模型 16
3.3.3經語速正規化之音素長度預估模型 17
3.3.4決策樹的建置 18
3.3.5使用語言參數產生決策樹分類 20
3.3.6加入韻律參數產生決策樹分類 21
3.4 音素能量位階 22
3.4.1音素能量位階之正規化 22
3.4.2四維正交基底展開 22
3.5音素能量位階之估計 23
3.5.1直接對音素ID去做分類 23
3.5.2決策樹的建置 23
3.5.3根據音素在音節之位置分四類並使用決策樹分類 23
3.5.4每種音素各自建立決策樹分類 25
3.6音素基頻軌跡 26
第四章 實驗結果與分析 27
4.1音素長度模型訓練之結果與分析 27
4.1.1音素長度正規化結果分析 27
4.1.2音素長度估計模型訓練結果/音素長度估計結果分析 35
4.2 音素能量估計結果與分析 37
4.2.1音素能量位階形狀分析 37
4.2.2音素能量形狀分析之固定音素長度 41
4.3音素基頻之分析 43
4.3.1音素之voiced比例 43
4.3.2統計voiced phone接著voiced phone的pitch狀況 45
4.3.3分析有聲音素接有聲之unvoiced比例 46
第五章 結論與未來展望 55
參考文獻 56
附錄一 57
附錄二 59
附錄三 61
附錄四 64

[1] WaveSurfer Homepage：www.speech.kth.se/wavesurfer/
[2] The HTK Book (for HTK version 3.4)
[3] Z. Sheng, J.-H. Tao, and D.-L. Jiang, “Chinese prosodic phrasing with extended features,” Proceedings of the IEEE ICASSP 2003, Vol. 1, pp. 492–495.
[4] C.-Y. Tseng, S.-H. Pin, Y.-L. Lee, H.-M. Wang, and Y.-C. Chen, “Fluent speech prosody: Framework and modeling,” Speech Commun. special issue on quantitative prosody modeling for natural speech description and generation, 46, 284–309 (2005).
[5] Chen-Yu Chiang, Sin-Horng Chen, Hsiu-Min and Yu, Yih-Ru Wang, “Unsupervised Joint Prosody Labeling and Modeling for Mandarin Speech,” J. Acoust. Soc. Am., vol. 125, No. 2, pp. 1164-1183, Feb, 2009.
[6] S.-H. Chen and Y.-R. Wang, “Vector quantization of pitch information in Mandarin speech,” IEEE Trans. Commun., vol. 38, no. 9, pp. 1317-1320, Sept. 1990.
[7] 賴佳鴻, ”跨語料庫之邊界模型對自動化切割的改善”, 2017
[8] 謝喬華, “考慮語速影響之漢語韻律模型建立 與語音合成之應用”, 2012
[9] T. Yoshimura, “Simultaneous modeling of phonetic and prosodic parameters, and characteristic conversion for HMM-based Text-ToSpeech systems’’ PhD dissertation, Nagoya Institute of Technology, 2002.