Research On Anti-noise Speaker Recognition Based On Non-reconstructed Compressive Sampling

With the development of computers and internet technology, speaker recognition as a kind of biological authentication,has got wide attention in the field of human-computer interaction.Speaker recognition has developed from the laboratory to practical application,and users’ demand for its accuracy,friendliness and robustnesss is higher.In real environment,the performance of speaker reconition is affected by many factors and one of the most important factors is environmental noise.Speech signals polluted by environmental noise,which make the matching degree between trainning data set and test data set down,cause the degradation of identificaition performance.On the other hand,with the development of cloud computing and big data,people have access to the increasing amount of information.In order to reduce the burden of transmission and processing data,signal compressed sensing technology arises. Speaker recognition under Nyquist sampling has got a large amount of data in order to ensure a high recognition rate, resulting in a waste of sampling resources, and compressive sensing theory can solve this problem. The thesis extracts feature parameters of measurement sequences after measurement matrix and undertakes research on robust speaker recognition based on compressed sensing,which is a part of national natural science foundation work of the instructor,and main works are as follows:(1)Studied the recognition system performance caused by compression ratio and frame length,and analyzed the measurement sequences after line ladder matrix and the influence of compression ratio and frame length on recognition rate.Recognition performance is almost the same as that of the traditional method when compression ratio is 1:2.Frame length is too short or long so that speaker recognition has obvious downtrend,and the frame length is about 20-30 ms.(2) Studied feature parameters and de-noising technology of recognition system based on compressed sensing in noisy environment.Firstly,the feature parameter CS-SSMFCC is proposed based on spectral subtraction,which improves robustness.Moreover,wavelet threshold de-noising is applied to the front-end system,which is better than spectral substraction.Under 15 db SNR,the recognition rate can teach more than 90%.(3)Proposed another key feature parameter-pitch frequency based on line ladder matrix,and then the two outputs of pitch information and Mel cepstrum coefficient are linear weighted processed to improve the recognition rate.The thesis uses both fixed weighted and dynamic weighted fusion method according to the weighted coeffient whether related to every testing voice or not.Experimental results show that recognition performance of dynamic weighted fusion is better.On this basis,a new robust fusion recognition system is proposed with wavelet soft threshold applied to the front-end,which enhances robustness.