A multilayer perceptron postprocessor to hidden Markov modeling for speech recognition

A novel neural network postprocessor for enhancing the classification capability of hidden Markov modeling for speech recognition is introduced. This postprocessor receives stimuli not from one but from all word HMMs and does not require segmentation of speech frames at the subword level. This postprocessor achieved 20% to 30% initial part error reduction on an HMM (hidden Markov model)-based isolated Chinese whole syllable speech recognition system, and can also be used for continuous speech recognition.<>