Calculation of Model Parameters Based on Vector Quantization in Speaker Identification Using Gaussian Mixture Models

The use of Gaussian Mixture Model (GMM) is most common in speaker identification. The most of the computational processing time in GMM is required to compute the likelihood of the test speech of the unknown speaker with consider to the speaker models in the database. The time required for speaker identification is depending to the feature vectors, their dimensionality and the number of speakers in the database. In this paper, we focused on optimizing the performance of Gaussian mixture (GMM) and adapted Gaussian mixture model (GMM-UBM) based speaker identification system and proposed a new approach for calculation of model parameters by using vector quantization (VQ) techniques to increase recognition accuracy and reduce the processing time. Our proposed modeling is based on forming clusters and assigning weights to them according to upon the number of mixtures used for modeling the speaker. The advantage of this method is in the reduction in computation time which depends upon how many mixtures are used for training the speaker model by a substantial value compared with approaches which use expectation maximization (EM) algorithm for computing the model parameters.

Keywords:

Speaker identification , Gaussian mixture model , EM algorithm , Vector quantization , Feature extraction