A neural predictive hidden Markov model for speaker recognition

An automatic speaker identification system is dereloped based on a neural predictive HMM speech recognition system. The neural predictive HMM utilises layered feed-forward neural networks to implement joint linear/nonlinear speech-frame prediction. A Markov chain is used to control changes in the network's weight parameters. The system exhibits high accuracy speaker recognition using very short speech utterances. The system was also shown to be robust against intraspeaker speech variations. The developed system is also well suited for efficient parallel VISI implementations enabling real time speaker identification performance.