文摘
The mismatch between the training and the testing environments greatly degrades the performance of speaker recognition. Although many robust techniques have been proposed, the mismatch problem is still a challenge for speaker recognition system. To solve this problem, we propose an optimized dictionary based sparse representation for robust speaker recognition. To this end, we first train a speech dictionary and a noise dictionary, and concatenate them for sparse representation; then design an optimization algorithm to reduce the mutual coherence between the two learned dictionaries; after that, utilize mixture k-means to model speaker corresponding to sparse feature; and finally, present a distance divergence to measure the similarity. Compared with the Mel-frequency cepstral coefficients based speaker recognition, our preliminary experiments show that the proposed recognition framework consistently improve the robustness in the mismatched condition.