Identifying the influence of transfer learning method in developing an end-to-end automatic speech recognition system with a low data level

Ensuring the best quality and performance of modern speech technologies, today, is possible based on widespread use machine learning methods. The idea this project to study implement an end-to-end system automatic recognition using methods, as well develop new mathematical models algorithms for solving problem agglutinative (Turkic) languages. Many research papers have shown that deep methods make it easier train systems approach. This method can also directly, is, without manual work with raw signals. Despite good quality, model has some drawbacks. These disadvantages are need a large amount data training. serious low-data languages, especially Turkic languages such Kazakh Azerbaijani. To solve problem, various needed apply. Some used belonging group same family (agglutinative languages). Method low-resource transfer learning, resources – multi-task learning. increase efficiency quickly associated limited resource, was model. helped fit trained dataset Azerbaijani dataset. Thereby, two language corpora were simultaneously. Conducted experiments show reduce symbol error rate, phoneme rate (PER), by 14.23 % compared baseline (DNN+HMM, WaveNet, CNC+LM). Therefore, realized be recognize other