Construction and Evaluation of a Speech Emotion Classifier using LSTM

The amount of video data has been increasing in recent years. Therefore, it is required to extract only important scenes from video data. Emotion of a person in a video can be used to find important scenes. Speech is one of the data useful to estimate emotions. If the accuracy of emotion classification using speech data is improved, it will be possible to estimate emotions even for videos in which a person does not clearly show facial expressions, and extract scenes in which a person strongly expresses emotions. In this study, we focus on speech data. We create a model using LSTM to classify speech utterances into appropriate emotions. We extract speech features as time series data from a dataset of speech utterances. Using these features, we train and evaluate the LSTM-based speech emotion classifier. Experiments are conducted in 5-fold cross-validation to evaluate the classification accuracy. As a result, the overall classification accuracy in this study was 24.5%.