Speaker identification using distributed vector quantization and Gaussian mixture models

Abstract

Speaker identification is the computing task of recognizing people's identity based on their voices. There are two main difficulties in this field. First is how to maintain the accuracy rate under large amount of training data. Second is how to reduce the processing time. Previous studies reported that Gaussian Mixture Model (GMM) for speaker identification appears to have many advantages. However, due to long processing time, this process does not always produce satisfying result in practice. Meanwhile, current mechanisms for hybrid production of speaker identification are directed more towards accuracy problems, not processing time optimization. This research focuses on constructing distributed data training on Vector Quantization (VQ) modeling to achieve an initial result. The decision tree approach is applied to obtain distributed training for VQ model. GMM classification process is then employed on the initial result to achieve a final result. The efficiency of the model is evaluated by computational time and accuracy rate compared to GMM baseline models. Experimental result shows that the hybrid distributed VQ/GMM model yields better accuracy. Besides, it gives 80% reduction in processing time and is 5 times faster compared to GMM baseline models. In conclusion, this research successfully improves the computational time and accuracy of the text-independent speaker identification system.