結合ANN預測、全域變異數匹配與真實軌跡挑選之基週軌跡產生方法

A Pitch-contour Generation Method Combining ANN Prediction, Global Variance Matching, and Real-contour Selection

基週軌跡（pitch contour）對於合成出高自然度的語音信號是相當的重要的，因此本論文研究提出了一種新的基週軌跡產生方法，此方法就是把類神經網路（artificial neural network, ANN）預測模組、全域變異數匹配（global-variance matching, GVM）與真實基週軌跡挑選（real contour selection, RCS）模組作結合，用以產生基週軌跡。在此，我們先分析出各個訓練音節的基週軌跡，然後使用離散餘弦轉換（discrete cosine transform, DCT）將各個基週軌跡轉換成對應的DCT係數之向量，然後就可拿各個訓練語句的DCT向量序列、及對應的語境參數去訓練ANN權重值與GVM參數。在基週軌跡產生的實驗中，我們以量測變異數比值（variance ratio, VR）來作為客觀評估的依據，由實驗結果得知，GVM與RCS模組有助於提升VR值；此外，主觀聽測實驗的結果顯示，ANN加GVM所產生的基週軌跡，其自然度比僅使用ANN模組的高，並且ANN加GVM加RCS的基週軌跡自然度，更高於ANN加GVM的。

Pitch contours are important for synthesizing highly natural speech signal. In this paper, we study a new pitch-contour generation method. The method proposed is to combine ANN prediction module with global-variance matching (GVM) and real contour selection (RCS) modules. Here, a syllable pitch contour is first analyzed and then transformed via discrete cosine transform (DCT) to a DCT-coefficient vector. Each sequence of DCT vectors analyzed from a training sentence plus contextual parameters are then used to train the ANN weights and GVM parameters. In pitch-contour generation experiments, we measure variance-ratio (VR) values for objective evaluations. The modules, GVM and RCS, are shown to be helpful to promote VR values. In addition, in subjective evaluation, the pitch-contour generation method, ANN + GVM, is shown to be more natural than the method, ANN only. Also, the method, ANN + GVM + RCS, is shown to be better than ANN + GVM.