论文部分内容阅读
为解决跨数据库语音情感识别领域中实验数据集特征不匹配的问题,提出一种基于时频原子的听觉注意特征提取模型.首先,为了提取频谱特征,引入听觉注意模型对多类情感特征进行有效的探测.然后,利用选择注意机制改进了提取的语谱图特征,其中包含的显著性信息与跨库识别性能有紧密联系.再引入Chirplet时频原子,通过形成的过完备原子库提高语谱图特征的信息量.来自多个数据库的样本具有多成分分布的特征,据此所提模型中的Chirplet扩大了特征向量在时频域上的尺度.实验结果显示,相比传统特征模型,所提方法性能有显著提升.此外,该方法在训练集和测试集来源不一致情况下具有更好的鲁棒性.
In order to solve the problem of mismatch of experimental datasets in cross-database speech emotion recognition, a time-frequency-based auditory attention feature extraction model is proposed.Firstly, in order to extract the spectral features, the auditory attention model is introduced to validate many kinds of affective features Then, the selective attention mechanism is used to improve the extracted spectral features, and the salient features contained in them are closely related to the performance of cross-database recognition. Then the Chirplet time-frequency atom is introduced to improve the spectrum through the over-complete atomic library The information of the feature of the graph.Multiple samples from multiple databases have the feature of multi-component distribution, and the Chirplet in the proposed model expands the scale of the feature vector in the time-frequency domain.Experimental results show that compared with the traditional feature model, In addition, this method has better robustness in the case of inconsistent sources of training set and test set.