Speech Recognition and Machine Translation Using Neural Networks

The application of deep recurrent neural networks LSTM (long short-term memory) for English-Russian translation from speech to text is studied. Deep neural network learning and validation was performed based on English audiobooks. The neural network learning dataset was compiled by splitting an English audiobook and Russian text into sentences, and extracting features from audio files. Algorithms for learning and validating a neural network are available as part of the TensorFlow machine learning library. The calculations were performed on a Linux server with two high-performance NVIDIA video cards. Two types of models were learned: models for translating text to text and models for translating speech to text. As a result of the work, it was revealed that models based on deep neural networks are effective for machine translation of a language.