Speaker Recognition by Hidden Markov Models and Neural Networks

摘要

As humans, we develop the ability to identify people by their voice at an earlyage. Getting computers to perform the same task has proven to be an interesting problem. Speaker recognition involves two applications, speaker identification and speaker verification. Both applications are examined in this effort. Two methods are employed to perform speaker recognition. The first is an enhancement of hidden Markov models. Rather than alter some part of the model itself, a single-layer perceptron is added to perform neural post-processing. The second solution is the novel application of an enhanced Feature Space Trajectory Neural Network to speaker recognition. The Feature Space Trajectory was developed for image processing for temporal recognition and has been demonstrated to outperform the hidden Markov model for some image sequence applications. Neural post-processing of hidden Markov models is shown to improve performance of both aspects of speaker recognition by increasing the identification rate from 70.23% to 88.44% and reducing the Equal Error Rate from 3.38% to 1.56%. In addition, a new method of cohort selection is implemented based on the structure of the single layer perceptron. Feasibility of using Feature Space Trajectory Neural Networks for speaker recognition is demonstrated. Favorable identification results of 65.52% are obtained when using a large training database. The FST configurations tested outperformed a comparable HMM system by 12-24%.