The Research Of Speech Emotion Recognition Based On CNNs

As a branch of computer science, artificial intelligence, essentially, not only simulates the process of reflecting and thinking of human being, but also surpasses human intelligence. Since artificial intelligence research is the study of human intelligence, it is necessary to focus on human emotion. And through the research of speech emotion recognition, the computer will be endowed with more humanization and pertinence in human-computer interaction.Deep learning is also a hot and difficult subject of the artificial intelligence research. Facing with the problem of abstract cognition, self- learning the feature of study object, one of the characteristics of deep leaning, have achieved great success in both of theory and application of artificial intelligence. It says that the deep learning is so far the most similar leaning method to the human brain. There are a number of common models of deep learning, such as auto-encoder, restricted boltzmann machine, convolutional neural networks, etc. Among them, the convolution neural networks have been applied in image recognition, speech recognition, advertisement recommendation system and so on, obtained a significant development.In this paper, after in-depth study of deep learning, we found that when when we use conventional convolution neural networks for speech emotion recognition, sub-sampled layer will achieve the effect of dimensionality reduction, but also may lead to the loss of the majority of useful features which may reduce the accuracy of speech recognition and classification of emotion.To deal with these problems, this paper will proposes a improved algor ithm based on the convolution neural networks. In this algorithm, we use the feature selection method instead of the simple process of the sub-sampling layer, which is called convolution feature selection neural networks. Meanwhile, it’s proved that the improved algorithm is better than the traditional convolution neural networks through experiments.