Abstract

The neural network-based approach to the classification of implicit emotional components in ordinary speech is considered. Mel-frequency cepstral coefficients were used as feature vectors, and the multilayer perceptron with one hidden layer was used as the classifier. It was shown that the neural-network system developed is able to classify these kinds of speech with the accuracy up to 99% and is not inferior to the human experts. Moreover, two model’s training approaches were suggested and tested, and the influence of the parameters for mel-frequency cepstral coefficients calculation on the resulting accuracies was studied. It was found that the personalized approach to training the classifier for each subject results in higher classification accuracy than the generalized one that is, using a mixed sample of multiple subjects. Optimal parameters for the mel-frequency cepstral coefficients calculations were found. The results of the study demonstrated high quality of the developed approach, and it can be applied to developing Brain-Computer interfaces based on inner speech patterns recognition, which will be addressed in further research.

中文翻译：

---

一种基于神经网络的普通语音内隐情感成分分类方法

摘要

考虑了基于神经网络的普通语音中隐含情绪成分的分类方法。将梅尔频率倒谱系数用作特征向量，并将具有一个隐藏层的多层感知器用作分类器。结果表明，所开发的神经网络系统能够以高达99％的准确度对这些语音进行分类，并且不逊色于人类专家。此外，提出并测试了两种模型的训练方法，并研究了梅尔频率倒谱系数计算参数对所得精度的影响。已经发现，针对训练每个主题的分类器的个性化方法比使用多个主题的混合样本的广义方法具有更高的分类精度。找到了梅尔频率倒谱系数计算的最佳参数。研究结果证明了所开发方法的高质量，并且可以将其应用于基于内部语音模式识别的脑机接口的开发，这将在进一步的研究中得到解决。