A Research Of Speaker Recognition Based On Sparse Representation Under Noisy Environment

As a technology of voiceprint identification,speaker recognition has unlimited prospect in the day of the application of pattern recognition developing rapidly,it is more convenient and lower cost than any other identification methods which is use individual biological characteristic to make a recognition.In recent years,the study of speaker recognition has attracted widespread attention.At present,the common model of speaker recognition model is Gaussian Background Mixture model,this model is based on the training of general background model,and its robustness is better than any other models,but this way has a large amount of calculation and the effect of recognition is unsatisfied.So,there are many people to make improvements based on this model.In recent years,the sparse representation algorithm has an astonishing performance in the field of signal processing,and it has been achieved a good effect in the images' recognition,processing and separation.In addition,letting the sparse representation as a classification algorithm to improve the speaker recognition system,introducing it into the matching recognition module and using the characteristic of sparse representation to solve the question in the speaker recognition system about the efficiency of recognition will drop down sharply when there are some noise interference.The main work of this paper includes the following contents:First of all,we put the algorithm of sparse representation into the speaker recognition model,using the classified characteristics of sparse representation to make an improvement of the model's recognition,and using the calculation of minimum standard reconstruction error to find the corresponding speaker.In the second place,we make a design of the dictionary form to meet demand of sparse representation algorithm.Using the most popular GMM mean super vectors as dictionary atoms.Aiming at the problem of super vectors' dimension is too large,we use the Fisher discriminant ratio to compare the classification performance of dictionary in each dimension,and make a rule to control the reduced dimension of dictionary,and add unit matrix of I into the dictionary to improve noise immunity of the system.Using the simulation can prove two things.The first thing is we can get better recognition efficiency when we mix the sparse representation into the model of speaker recognition.The second thing is the I-Fisher algorithm which is proposed in this paper can not only reduce the dimension of the dictionary,but also improve the recognition performance of the system.This recognition model is suitable for the test speech and training speech which was recorded in the same conditions,and it has a good effect of recognition.But if we want the recognition model can be fitted in any conditions of noise environment,we need to train multiple dictionaries,and there is a large amount of calculation.Thirdly,aiming at the problem of recognition rate decreased in different noise environment,we propose a new dictionary construction method based on sparse representation to solve the negative effect of noise.According to the principle of MCA morphological component analysis,we use clean speech to train the speaker dictionary,and we use the way of add noise dictionary to divide the sparse representation coefficient into two parts: the pure speech coefficient and the noise speech coefficient.After this,we make a calculation of reconstruction error to the part of clean speech coefficient,in order to recognize speaker under a condition without the effect of noise on recognition effect.In order to get the dictionary in which can be meet the design requirements,we use K-SVD dictionary learning method to train two dictionaries separately and put them together,and put the noise dictionary as a part of speaker dictionary into big dictionary to get coefficient by the decomposition of sparse representation.And we propose a method to make a same decomposition of noise test voice,in order to extract noises to update the noise dictionary.The paper use simulation to prove the algorithm can reduce the effect of noise recognition effectively when the testing speech and training speech in different conditions.The paper puts forward two recognition models of sparse representation based on the noise environment.We make an improvement of atomic construction to the first dictionary,using the experiment to test its applicable conditions of recognition.We also purpose a design project of the second dictionary based on the noise dictionary,and make a update to the noise dictionary.The two recognition models of sparse representation have achieved good recognition effect.