Research On The Robust Speaker Recognition Based On Sparse Representation

After decades of research and development, speaker recognition technology has got wide attention in because of its union advantages. In the research of speaker recognition, model is one of key technologies and has a great influence on the performance of the system. The theory of sparse representation has been widely employed in speaker recognition in recent years and improved the performance of the recognition system combined with Guassian mixture models. Nowadays, the main direction of speaker recognition research is to improve it’s robust.and to recognize quickly and exactly with intelligent terminal. This thesis undertakes a study on speaker recognition based on sparse representation and completes hereinafter works:(1)In order to decrease the demand of training speeches, this thesis proposes to use Guassian mixture model matrix instead of supervector to train redundant dictionary. Having done that, each person can own a redundant dictionary so that the calculated quantity in recognition process can be reduced.(2)This thesis contrasts the performance of exemplar dictionary and learning dictionary in clean and noisy environment respectively and has found that exemplar performs well in noisy environment. Meanwhile, this thesis proposes adding respective noise in dictionary to reduce environment differences. Simulation results show that this method greatly improve system’s recognition accuracy.(3)To address the problem of environment noise, this thesis proposes a universal compensation method applied to speaker recognition based on sparse representation. This method analyzes each feature vector members one by one to find the most corrupted one and remove it so that the correlation of the test utterances and training utterances will be enhanced. According to simulation results, this method has improved the system’s robustness to environment noise. The accuracy when SNR of white noise equals 15 dB can reach 96% which almost equals to the accuracy of recognizing a clean utterance.