Research On Tibetan Speech Recognition Technology Based On Recurrent Neural Network

As the most direct,natural and fast way of information interaction between people,voice plays an important role in people's daily life.Speech Recognition(SR)technology is one of the core technologies of human-computer interaction.Its goal is to let the computer convert the speech signal into the corresponding text or command through natural language understanding.The research of speech recognition has important theoretical significance and practical value in realizing voice dialing,voice navigation,indoor equipment control,voice document retrieval,and simple dictation data entry.With the development of neural network technology,cyclic neural networks that use context information to model the correlation of long-term speech have become the mainstream technology.As one of the ancient ethnic minorities in China,the Tibetans have a long history and splendid culture.Their language is an important part of Chinese culture and a bridge for cultural exchange between Tibetans and other brothers.In recent years,the speech recognition of Chinese and English has achieved fruitful results,and related methods and techniques are expected to become the pivot of Tibetan speech recognition.The recurrent neural network can mine the effective time series information in the input features,and enhance the distinguishing performance and expressive ability of the features.It has a greater advantage over traditional neural networks when dealing with continuous,context-sensitive tasks(such as speech recognition).Therefore,this paper studies the Tibetan speech recognition system based on recurrent neural network from the aspects of Tibetan text,acoustic characteristics and selection of modeling units.Firstly,by analyzing the characteristics of Tibetan characters,the Tibetan text preprocessing method,Tibetan character vector representation,and Tibetan language model are studied.Secondly,the speech data corresponding to the text is collected,and the Tibetan speech feature extraction method is studied.Combined with the characteristics of Tibetan speech,the phoneme is used as the modeling unit,and the acoustic model of Tibetan speech recognition is studied by establishing the probability mapping relationship between input speech and output sequence.Finally,the Tibetan speech recognition system was developed based on TensorFlow,and the experimental results of different super-parameters were evaluated by the label error rate and loss value.The experiment proves that the Tibetan speech recognition based on the cyclic neural network achieves good results under the test of closed corpus.