聲韻轉換及其在中文口語訓練之應用

Abstract

聲韻轉換旨在將來源語者的聲音藉由參數的調整轉換成具目標語者聲韻特性的聲音，其關鍵在於音長比例調整與音高比例調整兩項技術。在諧波正弦模型的分析合成平台下，利用動態時間校準求取各音框之縮放倍率，據以實現僅調整音框長度而不改變發音特性之時變音長比例調整。至於音高調整，則是分別以純量量化及向量量化兩種方式處理音高參數，設計最佳轉換機制以調整語音之聲調走向。本論文進一步將聲韻轉換技術應用於聽障中文口語的聲韻矯正，實驗結果顯示，依據動態時間校準的時變音長調整機制可以改善合成音質使其符合常人口語之說話速度，再配合基於向量量化的音高比例調整可提昇易於混淆的二聲及三聲聲調之辨識率。

Prosodic conversion is aimed to modify the speech signal of source speaker to match the prosodic properties of the target speaker. This conversion system consists of two subsystems, which are time-scale modification and pitch modification. Using a sinusoidal harmonic model, the former modifies the duration of source speech without changing the characteristics of vocal tract. This task was accomplished by adjusting the length of each frame in accordance with the dynamic time warping. Also proposed are two mapping functions that modify the characteristic features of pitch contour. One is based on scalar quantization and the other is based on vector quantization. Finally, we devise speaking aids to enhance the hearing-impaired speech by integrating spectral conversion and prosodic conversion. Experimental results indicate that the proposed approaches improve the identification accuracy of Mandarin tones, especially in tone2 and tone3.