CASIA OpenIR  > 毕业生  > 博士学位论文
汉语感叹句和疑问句的生成方法研究
其他题名Research on Generating Exclamantory and Question Speech in Mandarin
贾惠彬
2009-05-29
学位类型工学博士
中文摘要传统的语音合成多侧重于单一朗读语气的研究。为了进一步提高语音合成系统的表现力,扩展语音合成系统的应用范围。本文针对自然口语中常见的疑问句和感叹句进行了深入的研究,并且对韵律自动评价方法也进行了深入的探索。在分析了带有情态标记的感叹句和疑问句的韵律特点之后,通过构建新的韵律模板库和构建新的目标代价函数,在波形拼接合成系统下,实现了感叹句和疑问句的生成。并且,本文基于韵律转换的方法也模拟了感叹句和疑问句的生成。韵律评价是语音评价的一项重要研究内容,本文在分析了多个说话人间韵律可变性的基础之上,提出了自动韵律评价方法。这些研究成果对提高语音合成系统的表现力,促进合成系统的应用具有重要研究意义。具体来说,论文取得了以下研究成果: 感叹句和疑问句的韵律分析。本文借助于平行语料,分别对比分析了带有情态标记的感叹句与陈述句、四种疑问句与陈述句的韵律特点。从分析结果可以看出,在带有情态标记的感叹句中,情态标记常具有强重音,且强重音对相邻音节的时长和基频有一定的影响;在疑问句中,疑问标记词和疑问句语气词会提高相邻音节的音高和缩短相邻音节的时长。 基于波形拼接方法的感叹句和疑问句生成。借助于较小的感叹句和疑问句语料库,实现了一个感叹句和疑问句生成系统。该方法中,在韵律模板库中引入感叹句和疑问句的语气特征,基于该语气特征构建了新的目标代价函数。主观感知实验表明,该系统输出的感叹句和疑问句具有较高的自然度和较强烈的语气。 基于韵律转换方法的感叹句和疑问句生成。基于带有情态标记的感叹句和四种疑问句的韵律特点,通过引入该种感叹句和四种疑问句的语气特征,采用基于CART模型的韵律转换方法实现了感叹句和疑问句韵律生成,其基本思路为模拟感叹句中的强重音对相邻音节的影响,模拟疑问句的疑问标记词和疑问语气词对相邻音节的影响。主观评价试验结果表明了该方法的有效性与合理性。 韵律评价是语音评价的一项重要研究内容。本文从朗读风格的韵律模式入手,从音系层,即声调、基频走势和节律组织上分析了说话人间韵律的可变性,这种可变性为自动韵律质量评价带来了一定的困难。本文提出了一种自动韵律质量评价方法,通过计算待测语句与它的多个标准参考之间的声调、基频走势和节律组织的相似度,从而来评价待测语句的客观韵律质量。并且,通过对专家评分本质的分析,以及通过与多个分类与回归模型的比较,推断韵律质量评价问题可以认为是一个序回归问题。最后,通过在韵律评价库上进行测试,试验结果表明该方法取得了很好的人机评分相关度。
英文摘要Currently, Most Text-to-Speech system can synthesize speech only in a reading style, which greatly limited the application of TTS. To improve the expressiveness of TTS and to enlarge the application of TTS, this paper focuses on question and exclamantory speech in spoken language, and prosody evaluation is also investigated. Based on the analysis of the exclamantory speech with modal tags and question speech, a unit selection based method is proposed to generate exclamantory and question speech by constructing new prosody templates and target cost function. Besides, a prosody conversion based method is proposed to simulate excalmantory speech and question speech. Prosody evaluation is an important part of speech evaluation. In the paper, based on the analysis of prosody variability, a method of is proposed to automatically evaluate prosody quality. My achievements of this paper are as follows: In this paper, exclamantory speech with modal tags is analyzed and four types of question speech are also analyzed based on parallel speech. From the results, there are a few strong stresses in exclamantory speech with modal tags. Besides, question tag word and modal exclamation will increase neighbor’ F0 and decrease their duration. Generating exclamantory speech and question speech based on unit selection. Based on the analysis of question speech and exclamantory speech with modal tags, a new target cost function is constructed and new features are integrated into prosody templates. Experimental results show that the synthesized speech is of high quality and strong mood. Synthesizing exclamantory speech and question speech based on prosody conversion. Based on the prosodic analysis of question and exclamanatory speech, a new prosody conversion model is constructed with CART model to simulate the prosody features of exclamatory speech with modal words and question speech. Final perception and comparison experiments show that the models proposed can be used to synthesize the speech with high quality. Besides, it is showed that the method is valid. Prosody evaluation is an essential part of speech evaluation. The paper analyzes the prosodic variability among inter-speakers based on a speech database containing eight repetitions of sentences. For Mandarin, prosody variability can be analyzed from rhythm, intonation and tone variation, which is very difficult to automatically evaluate prosody quality. Based on these variations, the comparison between the tested an...
关键词语音合成 感叹句 疑问句 韵律评价 Speech Synthesis Exclamantory Speech Question Speech Prosody Evaluation
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6185
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
贾惠彬. 汉语感叹句和疑问句的生成方法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2009.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20061801462908(2334KB) 暂不开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[贾惠彬]的文章
百度学术
百度学术中相似的文章
[贾惠彬]的文章
必应学术
必应学术中相似的文章
[贾惠彬]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。