Speakerdependent automatic speech recognition in Buryat

Due to the dominance of Russian in education, formal communication, and mass media the usage of Buryat is gradually decreasing. Therefore, the problem of language preservation is of high concern for the native people. In modern conditions, in resolving this question significant consideration is given to speech technologies based on natural language processing, artificial intelligence and other methods to digitize the Buryat language and provide access to resources online. The analysis of current resources has found no Buryattargeted technologies that work with audio data. Therefore, the present study covers the development of speech recognition model prototype in 166 Buryat for automatic speech processing. The model uses the DeepSpeech2 architecture allowing to combine the functions of convolutional neural networks to extract and compress acoustic information relevant for recognition, as well as connections to identify time dependencies between phones in order to develop a model of a word in the Buryat language. The percentage of incorrectly recognized phones is 11%. This value significantly exceeds the values typical for modern models. The limitation of this model is the insufficient amount of training data and the dependence on the speaker's voice. Expanding the speech sample bank, annotating the data and optimizing the model can improve the accuracy of the system and its resistance to speech variability.