Research on Acoustic Modeling of Low Resource Uyghur Language Based on Transfer Learning

Abstract

Due to the shortage of Uyghur corpus resources, the performance of Uyghur acoustic models has been poor. Therefore, this paper proposes a Uyghur acoustic modeling method based on transfer learning, which uses the weight transfer method to transfer DNN hidden layers of large-scale English and Chinese acoustic models. In order to compare experiments, this paper builds a baseline system based on DNN-HMM. The training criteria are xEnt and MPE respectively, and the WER of the baseline system is 26.39% and 25.62% respectively. In this paper, 32 experiments were designed to find the best acoustic model. Finally, the optimal acoustic model WER was reduced to 18.75%. Through many experiments in this paper, it can be shown that the weight transfer method can effectively improve the performance of Uyghur acoustic model.