Efficient methods for rapid UBM training (RUT) for robust speaker verification

This study develops a computationally faster method for training background speaker models, with the goal of contributing robust speaker verification systems. The proposed method addresses the issue of speeding up the computational process of the system without hindering overall system performance. The method presented uses a sub-sampling scheme to allow for a more rapid snapshot of the speaker acoustic space, since acoustically similar adjacent frames are skipped to achieve smaller training material. This study also proves that for an effective UBM, it is better to use smaller amounts of data from a balance across a diverse speaker population, rather than blindly using all that is possible. The main disadvantage of using the entire set of data for training UBMs is that it is possible for biases to occur if some speakers have a significantly larger amount of data in the training pool. Therefore, it becomes important to find an efficient way to balance the number and distribution of the speaker population. The method presented in this study also addresses the impact of channel variability in speaker verification systems. The database used in our experiments is the National Institute of Standards and Technology (NIST) Speaker Recognition Evaluation (SRE) data from 04', 05', 06' and 08' data, and also a set of speakers from the FISHER database for Eigen-Channel Adaptation (for addressing channel variability). The results show a 70% significant improvement in computational speed of the system without any change in the robust speaker verification system performance. An analysis of the resulting acoustic space also provides evidence of the UBM training algorithm's advantages.