Research On Language Recognition Technology Based On Coordination Information

As the front end of Multi-speech recognition,language recognition plays an increasingly important role in international communication.The features in current language recognition system can be broadly divided into two types(i.e.acoustic feature and the coordination information of acoustic unit).The coordination information reflects the collocation relationships of acoustic units and is a kind of important information to describe the difference between different languages.The main method for obtaining coordination information is to convert the speech into an acoustic unit sequence by using a continuous speech recognition system,and then the acoustic units sequence is analyzed by statistical method.The flaw in the method is to build the continuous speech recognition system.Aiming this problem,two aspects is studied,one is to obtain coordination information from the acoustic model,which is established by the speech feature sequence.The other one is to use the acoustic unit discovery method to obtain the coordination information in the discovered acoustic units.On this basis,a language recognition system is built.The main work and innovations are as follows:Aiming at the problem that the language recognition system based on Gaussian Mixture Model-Universal Background Model(GMM-UBM)ignores the collocation relationships of acoustic units,a method based on Ergodic Hidden Markov Model(EHMM)is proposed to language recognition system.Each language need a corresponding EHMM to describe it,the states of EHMM correspond to acoustic units of language,the state-transitions reflecting the matching relations of acoustic units.The experimental results show that the proposed method gets a better result than language recognition based on GMM-UBM.The method of unsupervised acoustic units discovery is researched.Two approaches of acoustic units discovery are researched: parametric model and nonparametric model.The method of parametric model is to use GMM to discover acoustic units.The mixtures of GMM correspond to acoustic units,then clustering similar acoustic units to form a set of acoustic units.The nonparametric model is using nonparametric bayesian model to discover acoustic units.A Hierarchical Hidden Markov Model(HHMM)is used to model acoustic units,the states of topHMM correspond to acoustic units of speech.Then the set of acoustic units is obtained by unsupervised clustering using Hierarchical Dirichlet Processing(HDP).Finally the posterior probabilities of each acoustic units are obtained as the speech features,and the experimental results show that the posterior probabilities map reflects the acoustic segments distribution of speech.The language recognition method based on acoustic units discovery is proposed.The speech signal is converted into a posterior probabilities sequence of acoustic units based on acoustic units discovery.The method of N-gram joint posterior probability is used to compute the acoustic units' matching relations of each speech.The first step is to sum up and average the posterior probabilities of the frames that are considered to be within the same acoustic unit.The second step is to compute the joint probabilities for sequences of acoustic units.The final step is to sum up all matrices of the joint-posteriorgram,and convert the matrix into a vector as the feature of the speech.The method effectively avoids the sparseness of the counts matrixes.Then the i-Vector is used to reduce the dimension.Finally,language recognition is done with a Support Vector Machine(SVM).The experimental results show that the proposed method effectively avoids the dependency on the labeled speech and guaranteed the performance of language recognition system.