The article describes a research project in the field of Human-Computer Interaction, which focuses on the problem of automatic emotion recognition from speech. The aim of the project is to create a system that can recognize emotional states in the same way as humans to make the interaction between humans and digital machines more natural. The researchers used a modified recurrent neural network (RNN) architecture with long short-term memory (LSTM) to extract multiple temporal features from the audio signal. The experiment was done using an open dataset that includes eight different emotions: neutral, calm, happy, sad, angry, scared, disgust, and surprised. The accuracy of the model was tested using 40% of the dataset. The dataset used in the experiment is the TESS dataset, which includes 2800 audio data samples of two females speaking 200 target words with seven different emotions. The research used Python language and the Keras package, which is an open-source neural network library written in Python. The research found that the proposed deep learning - RNN architecture showed an outstanding performance on various problems.