普通话双模态情感语音数据库的设计与评价
2018-07-27任国凤张雪英李东闫建政
任国凤 张雪英 李东 闫建政



摘 要: 针对包含发音动作参数和情感语音的双模态汉语普通话数据库非常匮乏的问题,设计包含中性、高兴、愤怒及悲伤4种情感的普通话语音库。该语音库由10名被试录制的1 440段音频及发音动作数据组成,文本长度有双音节词和句子两种类型。为了确保该数据库的有效性,邀请普通话较好、听力正常的10名评价者组成评价小组,对数据库内所有音频文件进行评价。根据评价小组评价结果结合发音动作数据的稳定性进行筛选,得到语音质量较好、发音动作参数稳定的双模态情感语音数据库。该数据库可用于开展情感语音的发音动作研究,进而单独或联合作为情感语音识别算法的样本数据,对情感语音识别率的提高具有积极的作用。
关键词: 数据库; 情感语音; 发音动作参数; 汉语普通话; 信号处理; 普通话语音库
中图分类号: TN912.3?34 文献标识码: A 文章编号: 1004?373X(2018)14?0182?05
Design and evaluation of Mandarin bi?modal emotion speech database
REN Guofeng1,2, ZHANG Xueying1, LI Dong1, YAN Jianzheng1
(1. School of Information Engineering, Taiyuan University of Technology, Taiyuan 030600, China;
2. Department of Electronics, Xinzhou Teachers University, Xinzhou 034000, China)
Abstract: In allusion to the problem of lack of bi?modal Chinese Mandarin database containing pronunciation action parameters and emotional speech, a Mandarin speech corpus that includes four emotions of neutrality, happiness, anger and sadness is designed. The speech corpus is composed of 1 440 segments of audio and pronunciation action data recorded by 10 subjects, and the textual length includes bi?syllable word and sentence. To ensure the validity of the database, 10 evaluators with good Mandarin and normal hearing are invited to constitute the evaluation group, so as to evaluate all the audio files in the database. According to the evaluation results of the evaluation group and in combination with the stability of pronunciation action data, the audio files are screened to obtain the bi?modal emotion speech database with good audio quality and stable pronunciation action parameters. The database can be used to conduct the pronunciation action research of emotional speech, and solely or jointly taken as the sample data of the emotional speech recognition algorithm, which has a positive function for improvement of the emotional speech recognition rate.
Keywords: database; emotional speech; pronunciation action parameter; Chinese Mandarin; signal processing; Mandarin speech corpus
0 引 言
情感语音由于包含了丰富的情感信息,在人们日常交流中起着非常重要的作用,因而对其开展研究有助于进一步了解人们交流时的心理状态[1?2]。同时,伴随着人机交互智能技术的快速发展,情感识别能力作为度量人机交互智能化水平的重要依据,受到众多研究者的青睐。而情感语音识别的研究离不开优质的情感语音数据库的支撑。由于情感认知的复杂性、数据采集的困难性、录制标准的不统一性、情感分类目的和方法的差异性等,导致目前大多数研究机构都是按照各自的科研需求建立相应的情感语音数据库,而无法构建相对统一的情感语音库[3?4]。……
