基于规则的汉语情感语音系统的设计与实现

Design and implementation of rule-based emotional speech synthesis system

汉语的一个特点是存在大量的同音字和同音词。利用这一点,建立了一个基于音节的语音库,它覆盖了所有汉语单字,收录了三万余个汉语高频词。同时,考虑到相同的字处于词头和词的中间时发音有所不同这个特点,语音库中还加入了这两种情况下的单字的发音,提高了合成语音的自然度。最终得到一个大小为550M左右的包含合成所需基本词汇量的语音库。接着,采用波形拼接的方法,我们得到自然流畅的中性合成语音。在对预先录制的语音样本研究的基础上,总结出悲伤、喜悦这两种最基本情感的语音参数特点。最后,在中性合成语音上进行基频曲线特征、时长、信号平均强度等相关语音参数的调节,从而得出带有悲、喜情感的语音。听者测试实验表明,合成出的语音情感表现明显,证明了这个系统能以较小的容量代价实现基本情感语音合成。

In mandarin Chinese, one syllable may correspond to several single words or phrases. Employing this feature, a corpus was constructed, including all Mandarin single words and around 3000 frequently-used phrases. Meanwhile, pronunciations for the same word appearing in different places in phrases were also covered to improve the naturalness of the synthesized speech. Next, after analyzing pre-recorded emotional speech samples we concluded some acoustic features associated closely with two basic emotions, happiness and sadness. By imposing these rules on neutral speeches generated by the system, happy and sad synthesized speeches were obtained. Auditory test demonstrated the effectiveness of the approach.