The Research Of Uyghur Acoustic Model Based On Deep Neural Network

With the improving of the computer capacity in recent years, the Deep Neural Networks(DNN) technology has been deeply investigated. Deep Belief Networks(DBN) as one of the representative DNNs, successfully applied to speech recognition by a large number of scholars, and significantly improve the performance of the recognition system. At the same time, as an important part of multilingual information construction in Xinjiang, the Uyghur language for large vocabulary continuous speech recognition technology is of great research significance. Based on this background, in this paper, the DNN technology for more in-depth research, and focuses on how to use DBN technology in building Uyghur acoustic models. Finally, this paper verify effectiveness of this technology through experiments.The core aim of this paper is using DNN technology for optimization of Uyghur acoustic models to enhance the performance of the baseline recognition system.Around the core goal, this paper’s main study works as follows:(1)Researched Uyghur large vocabulary continuous speech recognition system,and focused on mastering the principle of acoustic model and modeling method;(2)Studied the DDN technology, emphatically grasped the Pre-training technology and Fine-tuning technology, understood the meaning of each parameter in DNN and DBN. On this basis, this paper studied how to build the Uyghur acoustic model based on DNN and DBN;(3)Studied Speaker Adaptation technology based on MLLR and MAP, and used this technology to optimize Uyghur acoustic model;(4)Conducted Uyghur acoustic modeling experiments based on DBN and SAT.The experimental results showed that the acoustic model based on DBN improved the baseline system’s recognition rate of 3.4%.