From artificial neural network inversion to hidden Markov model inversion: application to robust speech recognition

Abstract

The gradient based hidden Markov model (HMM) inversion algorithm is studied and applied to robust speech recognition tasks under general types of mismatched conditions. It stems from the gradient-based inversion algorithm of an artificial neural network (ANN) by viewing an HMM as a special type of ANNs. The HMM inversion has a conceptual duality to HMM training just as ANN inversion does to ANN training. The forward training of an HMM, based on either the Baum-Welch reestimation or gradient method, finds the model parameters /spl lambda/ to optimize some criteria (e.g., maximum likelihood, maximum mutual information, and mean squared error) with given speech inputs s. On the other hand, the inversion of an HMM finds speech inputs s that optimize some criterion with given model parameters /spl lambda/. The performance of the proposed gradient based HMM inversion for noisy speech recognition under additive noise corruption and microphone mismatch conditions is compared with the robust Baum-Welch HMM inversion technique along with other noisy speech recognition technique, i.e., the robust MINIMAX classification technique.