Speaker Verification Method Based on Two-Layer GMM-UBM Model in the Complex Environment

摘要

In order to improve speaker verification accuracy in the complex environment, a two-layer Gaussian mixture model-universal background model (GMM-UBM) model based on speaker verification method is proposed. For different layer, a GMM-UBM model was trained by different combination of speech features. The voice data of 3 days (36 h) were recorded from the complex environment, and the collected data was manually segmented into four classes: quiet, noise, target speaker and other speaker. Not only the segment data can be used to train GMM-UBM model, but also it can provide a criterion to assess the effectiveness of the model. The results show that the highest recall for the second and third day were 0.75 and 0.74 respectively, and the corresponding specificity were 0.29 and 0.19, which indicates the proposed GMM-UBM model is viable to verify the target speaker in the complex environment.