Research And Implementation Of Tibetan Language Model Based On RNN

With the rapid spread of the Internet and the rapid update of information,AI has become an important direction for the future development of science and technology.Speech recognition is an important branch of artificial intelligence research.Its purpose is to enable machines and people to communicate with each other through voice to realize human-computer interaction.At present,speech recognition has achieved a high recognition rate in large languages such as English and Chinese,but there is a relatively lack of research on small languages such as Tibetan.The language model is an important module in speech recognition,and it is also the main form of language factual relationship,which greatly affects the final effect of the speech recognition system.In addition to speech recognition,language models are also widely used in machine translation,automatic word segmentation,and syntactic analysis.This paper mainly studies the language model based on Recurrent Neural Network(RNN)and the traditional N-gram statistical language model,constructs relevant Tibetan language models and tests the performance of the model.By comparing parameters and adding optimization methods,the experiment compares the confusion of the two.The purpose is to obtain a Tibetan language model with better recognition performance,so that in the subsequent Tibetan speech recognition system,the acoustic model can be combined to obtain a more accurate recognition rate.The traditional N-gram language model is a shallow model.As the amount of data increases and the complexity of the data structure increases,the data will be sparse and so on,and its modeling ability will also decrease.The RNN is deeper,which has better learning and modeling capabilities than the N-gram model.In this study,by changing the number of hidden layer neurons in the RNN Tibetan language model,adding class layer acceleration operations in the output layer,and using context word vector features and LSTM training,the standard language model caused by gradient disappearance cannot effectively obtain long-distance constraints.The experimental results show that the optimized Tibetan RNN language model performs better than the traditional N-gram language model,but the training time is relatively long and the process is complex.