INFORMATION TECHNOLOGIES OF NEURAL NETWORK SPEECH RECOGNITION IN REAL-TIME

Relevance. Nowadays, it is relevant to study the basic means of audio signal processing, mainly from the point of view of sound classification and approaches to their improvement. The general characteristics of sound signals are considered, followed by a description of the time-frequency images for sound and the attributes useful for classification are reviewed. Human hearing is an incredible tool that gives us a lot of information about the world around us. We easily catch the sounds of birds, the sounds of cars at a distance, and even complex musical compositions. The subject of the study in the article is the human auditory system, which is able to process all this information, analyzing and grouping different sounds. This process is known as auditory scene analysis. Applications such as speech recognition, music transcription, and multimedia data retrieval can be greatly improved by separating and classifying sound sources. Digital audio signal processing has a number of important applications, such as audio compression, sound effect synthesis, and sound classification. Sound classification is becoming increasingly important as more and more multimedia content is created. This is especially useful when it comes to searching through audiovisual materials, as listening to audio clips can be a more efficient way to navigate than watching video scenes. Sound classification can also be used as an interface for audio compression, as different types of sounds, such as music and speech, require different compression methods. The purpose of this work is to explore approaches to building neural network speech recognition systems. Real-time speech recognition has become an incredibly useful tool for solving a variety of problems in different areas of life. Many companies now offer dictation software that allows people to create search queries or dictate emails using voice commands. It is appropriate to consider neural network speech recognition, in particular, Ukrainian. One of the biggest problems faced by the analysis of Ukrainian speech is the limited number of models available for recognition. While there are many models for English, there are very few for Ukrainian. In general, the potential benefits of sound processing and speech recognition are obvious, and it is quite likely that we will continue to see new developments in these areas in the future. Neural networks are described, the principle of their operation and methods of audio recognition using them. The following results were obtained: the audio signal, its representation, statistical and physical methods of working with it were studied. Conclusion. Effective models for correct speech recognition and toolkits for model training were found.