【作者】 许洁萍； 吕士楠； 闫俐；

【Author】 XU Jie-PingP1+P LU Shi-NanP2 PYAN LiP1P P1P(Department of computer, School of information, Renmin University, 100872, China) P2P(Institute of Acoustics, Academia Sinica, Beijing, 100080, China)

【机构】 中国人民大学信息学院计算机系； 中国科学院声学研究所；

【摘要】 目前,语音合成技术已经从人类的“理想”、“构建和谐社会”必不可少的手段,具体化为社会生活中方方面面的应用。对于汉语语音合成技术来说,是不是已经接近或达到了应用目标,是本文探讨的主题。文中简要报告了2004年在中国人民大学信息学院进行的汉语语音合成系统的评估实验结果,以此分析了汉语语音合成系统的可用性。评估采用主观印象评测(MOS)方法,选择了当前有代表性的五个汉语语音合成系统;从模块诊断评估和整体自然度评估评两个大的方面,测试了音节缺陷度、句子语义理解缺陷度及语句自然度测试和短文自然度四个方面。评估结果表明:目前的汉语语音合成系统虽然在儿化音等语音现象上还存在一定的问题,但在句子层面上都有能力躲避常见的语义陷阱;由于合成系统都是采用的拼接合成技术,合成系统的自然度大体都达到4分上下的水平,初步达到实用要求,都可以进入语音产品市场。更多还原

【Abstract】 At present, Chinese Text-to-Speech (TTS) system has been applied in many areas and affects people’s life. How about the performance of TTS is and what problem has been in TTS, it is our research aim. We reported the evaluation results for Chinese mandarin TTS system in Renmin University in 2004 briefly and studied the usability of Chinese TTS system at present. During testing, MOS (Mean Opinion Score) was used. Five Chinese TTS systems were chose. Considering of the modular diagnostic evaluation and the global naturalness performance, the testing contents included four items: the articulation of syllable, the intelligibility of sentence, the naturalness of sentence and the naturalness of paragraph. During the articulation of syllable testing, listeners determine the specified syllable to be good or not good. If we define that the inaccurate words are error ratios higher 80% and the contentious words are error ratios from 20% to 80%. The percent of inaccurate and contentious syllables are 23.5% respectively. These Results of articulation of syllable showed that there were defects in synthesis retroflex syllables for mandarin TTS systems. Considering of the effect of defective syllables for sentence, we define four grades for intelligibility of sentence: “fatal defective”, “some defective”, “light defective” and “no defective”. The fatal defective should be avoided. The testing results of intelligibility of sentence showed that the fatal defective is lower 20%.This results proved that TTS systems for mandarin have ability to avoid the familiar semantic hook at present. The naturalness is a key point for synthetic speech whether to be usability. We adopted the definition of five-score. Score 5 is the most natural speech just like the broadcasting speech, score 4.5 can be popularize in market. The testing result of naturalness of sentence is that the average of mean of every listener’s results is 4.17 and its average of standard deviation is 0.31. The average of mean for naturalness of paragraph for five systems is 4.2 and its average of standard deviation is 0.4. These results showed that the naturalness of sentences and paragraphs for five systems has achieved a higher level, which is between score 4 (good) and score 4.5 (to be popularized). So the TTS systems of mandarin can be applied in preliminary. The testing results are consistent with the info feedback from market.更多还原