Voice Based Emotion Recognition with Convolutional Neural Networks for Companion Robots

In order to obtain emotional-related response from robots, computers and other intelligent machines, the first and decisive step is accurate emotion recognition. This paper presents the implementation of this function with the deep learning model of Convolutional Neural Networks (CNN). The architecture wis an adaptation of an image processing CNN, programmed in Python using Keras model-level library and TensorFlow backend. The theoretical background that lays the foundation of the classification of emotions based on voice parameters is briefly presented. According to the obtained results, the model achieves the mean accuracy of 71.33% for six emotions (happiness, fear, sadness, disgust, anger, surprise), which is comparable with performances reported in scientific literature. The original contributions of the paper are: the adaptation of the deep learning model for processing the audio files, the training of the CNN with a set of recordings in Romanian language and an experimental software environment for generating test files. Key-words: Voice Recognition, Emotion Recognition Convolutional Neural Networks, Companion Robots, pet robots.