A Comparitive Survey of ANN and Hybrid HMM/ANN Architectures for Robust Speech Recognition

This paper proposes two hybrid connectionist structural acoustical models for robust context independent phone like and word like units for speaker-independent recognition system. Such structure combines strength of Hidden Markov Models (HMM) in modeling stochastic sequences and the non-linear classification capability of Artificial Neural Networks (ANN). Two kinds of Neural Networks (NN) are investigated: Multilayer Perceptron (MLP) and Elman Recurrent Neural Networks (RNN). The hybrid connectionist-HMM systems use discriminatively trained NN to estimate the a posteriori probability distribution among subword units given the acoustic observations. We efficiently tested the performance of the conceived systems using the TIMIT database in clean and noisy environments with two perceptually motivated features: MFCC and PLP. Finally, the robustness of the systems is evaluated by using a new preprocessing stage for denoising based on wavelet transform. A significant improvement in performance is obtained with the proposed method.