朝鲜语文语转换及其拼接代价的研究

拼接代价函数是决定合成音自然度的重要因素之一。针对拼接代价函数,以往的研究只考虑拼接点处的特征向量是否相等,而没涉及到特征向量的一阶连续性。该文研究并实现了以三音子做为基元的朝鲜语文语转换系统,并研究了基元动态特征对拼接的影响,设计了能反映特征连续性的一种新的拼接代价函数,利用两个基元在拼接点处的特征向量及其一阶差分值计算拼接代价。实验结果表明,基于动态特征连续性的拼接代价函数可以较好地改善拼接处的频谱连续性,有效地提高了朝鲜语合成语音的自然度。

Korean text-to-speech and concatenation cost function

The concatenation cost function is a key factor affecting the naturalness of synthesized speech.Previous research on concatenation cost functions has only discussed whether the characteristic vectors are the same at the concatenation point,but had not considered the first-order continuity.This paper describes a Korean text-to-speech(TTS) system based on triphones and the influence of the dynamic features of the triphones on concatenation.A concatenation cost function was developed which incorporates dynamic continuity by calculating the cost vector using the characteristic vectors and their first derivatives at the concatenation point.Test results show that the concatenation cost function improves continuity at the unit boundaries to substantially improve the naturalness of the Korean text-to-speech system.