華語韻律移植的改良

Improvement of Prosody Transplant for Mandarin Chinese

江蕙如(Hui-ju Chiang)

指導教授 ： 張智星

國立清華大學/電機資訊學院/資訊工程學系/碩士(2009年)

摘要

本論文主要目的是改良華語韻律移植的系統，將原始語音的韻律參數：聲調、長度及音量轉換成目標語音。不僅保有原始語音的音色，更著重於自然度與清晰度。
我們先對整句語句使用UPDUDP （unbroken pitch determination using dynamic programming）進行音高追蹤。首先對整個語句取出平均幅度差函數法的（average magnitude difference function, AMDF）時域特徵，再以動態規劃（dynamic programming, DP）擷取出連續而不中斷的音高特徵曲線。同時我們也用維特比譯碼器（viterbi decoding）進行強制對位（forced alignment）切出音素。
求取基週標位（pitch mark）時，則是採用基於動態規劃方法來求取基週標位。首先對於非氣音部分語音，先決定波峰或波谷做為基週標位的選取方式，再對語音中每個固定範圍內選取三個候選的基週標位後，計算其State probability與Transition probability，最後以動態規劃方式求取最佳的基週標位。
接下來取出非氣音部分的韻律參數進行韻律移植，合成方式是在時域上利用基週同步疊加法（pitch synchronous overlap and add, PSOLA）調整基週頻率以達成聲調轉換；用波形相似性疊加法（waveform similarity overlap and add, WSOLA）調整音長；再將原始語音與目標語音的每個基週標位到下一個基週標位的間隔當成音框大小（frame size）求取其音量矩陣，再利用音量矩陣調整原始語音的音量以達成音量轉換。

關鍵字

韻律移植

並列摘要

The purpose of this study was to show that an enhanced prosody transplantation system for Mandarin Chinese not only achieve pitch conversion, duration conversion, and energy conversion, but also upgrade naturalness and eliminate distortions.
UPDUDP (unbroken pitch determination using dynamic programming) is first applied to each utterance for pitch tracking which extracts an unbroken pitch contour from a given utterance based on time-domain acoustic feature of AMDF (average magnitude difference function) and DP (dynamic programming). Utterances are then segmented into phonemes using Viterbi decoding for forced alignment.
A DP-based pitch marking method is utilized for detecting pitch marks in a reliable manner. First we select either peaks (local maxima) or valleys (local minima) for pitch mark candidates according to its similarity to an estimated pitch curve. Based on the candidates, we define state and transition probabilities and then employ DP to find the most likely pitch marks.
The voiced characteristics of each utterance are then extracted to perform prosody transplantation. The PSOLA (pitch synchronous overlap and add) technique in time domain is used to adjust the fundamental frequency to achieve pitch conversion. WSOLA (waveform similarity overlap and add) is employed to adjust duration. Frame size is set to be the interval between one pitch mark and the next pitch mark of source wave and target wave. The volume for each frame is thus computed and linear mapping is executed. Energy conversion is then achieved by adjusting the source energy to the target energy.

並列關鍵字

UPDUDP ； PSOLA ； WSOLA

參考文獻

[8] 詹詩涵, 基於音高調節之歌聲合成系統., 國立清華大學資訊系統及應用研究所, 碩士論文, 2006.

[7] 石文俐, 中文語音合成之韻律產生器的改良與研究., 國立清華大學資訊工

[2] 江克敬, 華語韻律轉換之研究與實作., 國立清華大學資訊工程所, 碩士論

[12] Jiang-Chun Chen, J.-S. Roger Jang, "TRUES: Tone Recognition Using Extended Segments", ACM Transactions on Asian Language Information Processing, 2008.

[15] Cheng-Yuan Lin and J.-S. Roger Jang, "A Two-Phase Pitch Marking Method for TD-PSOLA Synthesis", GESTS International Transaction on Speech Science and Engineering, No. 2, Vol. 1, pp. 211-221, Dec 2004.