Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-08-13 Cooperative journals: 《计算机应用研究》
Abstract: In view of the low resource of endangered languages, the establishment and research of end-to-end speech recognition model can explore new ways for the protection and transmission of endangered languages. this paper combined dynamic bi-directional long short-term memory network and connectionist temporal classification model into an end-to-end speech recognition model. When performing phoneme-level recognition training, the batch size of the data passed into the model can be adaptively adjusted according to the training model, which not only speeds up the convergence but also improves the generalization of the model. By adjusting the hierarchy of the deep neural network and extracting different phonetic features for model comparison, the experimental results show that both the endangered languages - Lvsu and Tujia have good recognition results.