融合多种语言的语音情感识别

Speech emotion recognition integrating multiple languages

由于语言的差异,提高跨语言情感数据库识别语音情感的准确度,仍然是一项难题。该文针对语言差异这一难题,融合了语音情感识别技术和自然语言处理技术。该文选取Berlin语音情感数据库和CASIA语音情感数据库,从两个数据库中分别挑选200条语音,选用开源API下的Google Speech,实现语音文本的转化。使用机器翻译方法,将语言转化为文本,统一翻译成中文。利用自然语言处理的词法分析、句法分析、LSA的关键词提取算法,提取出表达情感的关键词。对于被提取出来的关键词,使用SpeechLib工具包将提取过特征值的文本转化成语音,提取MFCC特征,构建DNN+BLSTM模型,实现语音情感的分类。实验结果表明,文中使用的方法未加权平均召回率(UAR)和加权平均召回率(WAR)分别为48.22%和56.5%,相比其他方法,UAR和WAR分别提高了4%和8%。

Aiming at the problem of language differences,this paper selected Berlin voice emotion database and CASIA voice emotion database,selected 200 voices respectively,and selected Google speech under the open source API to realize the transformation of voice text. Use machine translation to translate the text into Chinese. Using the lexical analysis,syntactic analysis and LSA keyword extraction algorithm of natural language processing,the keywords expressing emotion were extracted. The extracted keywords were transformed into speech using speechlib toolkit to extract MFCC features,built DNN +BLSTM model and completed classification. The experimental results showed that the unweighted average recall(UAR)and weighted average recall(WAR)of the method used in this paper were 48.22% and56.5% respectively.