基于特征迁移学习方法的跨库语音情感识别

Cross-corpus speech emotion recognition based on a feature transfer learning method

在实际语音情感识别系统中,训练语音和测试语音往往来自不同的语料库,识别率下降显著。针对这一问题,该文提出一种有效的基于特征迁移学习的跨库语音情感识别方法。引入最大均值差异（maximum mean discrepancy,MMD）来描述不同数据库情感特征分布之间的相似度,并通过最大均值差异嵌入（maximum mean discrepancy embedding,MMDE）算法及特征降维算法来寻找二者之间的邻近低维特征空间,并在此低维空间中训练得到情感分类器用于情感识别。同时为了更好地保证情感信息的类别区分度,进一步引入半监督判别分析（semi-supervised discriminant analysis,SDA）方法用于特征降维。最后在2个经典语音情感数据库上对提出的方法进行实验评价,实验结果表明：提出的方法可以有效提高跨库条件下的语音情感识别率。

Speech emotion recognition systems offen use training data and testing data from different corpora, so the recognition rates decrease drastically. This paper presents a feature transfer learning method for cross-corpora speech emotion recognition. The maximum mean discrepancy （MMD） is used to describe the similarities between the emotional feature distributions of the different corpora, then the latent close low dimensional feature space is obtained via the maximum mean discrepancy embedding （MMDE） and dimension reduction algorithms, with the classifiers then trained in this space for emotion recognition. A semi supervised discriminative analysis （SDA） algorithm is further used for dimension reduction to better ensure the class discrimination of the emotional features. Tests on two popular speech emotion datasets demonstrate that this method efficiently improves the recognition rates for cross-corpora speech emotion recognition.