用于跨库语音情感识别的DBN特征融合方法

发布时间：2019-01-22 10:39

【摘要】：跨数据库语音情感识别中,将不同尺度上提取的情感特征进行融合是目前的技术难点。本文利用深度学习领域的深度信念模型,提出了基于深度信念网络的特征层融合方法。将语音频谱图中隐含的情感信息作为图像特征,与传统情感特征融合。研究解决了跨数据库语音情感识别中,将不同尺度上提取的情感特征进行融合的技术难点。利用STB/Itti模型对语谱图进行分析,从颜色、亮度、方向三个角度出发,提取了新的语谱图特征;然后研究改进的DBN网络模型并对传统声学特征与新提取的语谱图特征进行了特征层融合,增强了特征子集的尺度,提升了情感表征能力。通过在ABC数据库和多个中文数据库上的实验验证,特征融合后的新特征子集相比传统的语音情感特征,其跨数据库识别结果获得了明显提升。
[Abstract]:In cross-database speech emotion recognition, it is a difficult technology to fuse emotion features extracted from different scales. In this paper, a method of feature level fusion based on deep belief network is proposed by using the deep belief model in the field of deep learning. The emotional information implied in the speech spectrum is taken as the image feature and fused with the traditional emotional feature. This paper studies and solves the technical difficulties of fusion of emotion features extracted from different scales in cross-database speech emotion recognition. The STB/Itti model is used to analyze the spectrogram, and a new feature is extracted from three aspects: color, brightness and direction. Then the improved DBN network model is studied and the traditional acoustic features and the newly extracted spectral features are fused into the feature layer to enhance the scale of the feature subset and enhance the ability of emotional representation. The experimental results on ABC database and Chinese database show that the new feature subset after feature fusion is better than the traditional speech emotion feature, and the result of cross-database recognition is obviously improved.
【作者单位】：东南大学水声信号处理教育部重点实验室;烟台大学计算机与控制工程学院;
【基金】：国家自然科学基金项目(61673108,61231002,61375028) 山东省自然科学基金(ZR2014FQ016)
【分类号】：TN912.34

【相似文献】