Enhanced Language Model Personalization Based on Deep Belief Network

Since every mobile device is owned primarily by an individual person, it is possible to have a customized recognizer that well matches the qualities of that particular user. In personalisation of language models the need of much huge corpora(database) and massive amount of errors in the previous technologies makes them outdated and also they show poor performance during their execution. The previous technologies includes N gram based language model and recurrent neural network language model(RNNLM). Although projects on acoustic model personalisation has undergone lot of research over many years, but much less work has been reported on linguistic model personalisation. The primary section is the speaker recognition, which is done using MFCC(Mel frequency cepstrual coefficient processor). In the secondary section we will convert the speech input to text with predictive actions, like when we give a voice input to the system it will detect the word and convert it into the personalized text. This is called personalized, because the neural network were trained with the database of the specific user who is supposed to use the the system individually. For this, the data base will be created first with his personal slang. There might be grammatical errors posed by the individual user whilst training the neural network, however the same shall be used to create the database without correction along with other common words used by the individual user, so that it can generate the personalized language while he is talking to the system. Also it will display the next word prediction and the same will be taken from the personalized database. During the work of this thesis, we tend to propose improved customized model by using deep belief network. Here the deep belief network (DBN) shows higher performance than the N gram based models, recurrent neural network and conventional neural network technologies. Presumptively the deep belief network needs only less corpora and shows higher end in word error rate and running time. Also it is possible to have multiple word prediction with more accuracy using deep belief network, whereas with recurrent neural network it is possible only to have single prediction.