An algorithm for efficient speaker identification using reference speaker model based two-layer structure

To improve the GMM-UBM based speaker identification system’s computation efficiency, a fast algorithm using reference speaker model based two-layer structure is proposed. Vector quantization is used to train the reference speaker models using target speaker models. The deviations between one speaker and reference speaker models are used to model the speaker’s acoustic characteristics. The correlation between the deviations’ vectors is used to evaluate the similarity degree between the identified speech and target speakers and to prune those target speakers with lower similarity degree. Experimental results proved that on the database containing 5,200 target speakers and 1,000 out-of-set imposters, the proposed algorithm improved the identification efficiency by 12.5 times with performance degradation of only 0.3%.