Artificial Neural Network particularly Multi-Layer Perceptrons (MLPs) estimate the HMM state posterior probabilities was proposed by Bourlard et al, [13] which rely on a probabilistic interpretation of the MLPs outputs. Each output unit of MLP is trained to perform a nonparametric estimate of posterior probability of a left-toright CDHMM state given the acoustic observations. This represents a fundamental class of hybrid models. MLPs are used to estimate the state emission probabilities required in HMM as shown in Fig 3. The MLPs in this work consists of M input and N output nodes separated by a number of hidden nodes. The output of each layer forms the input of the consecutive layer. Each output node represents one symbol class corresponding to the phonetic units. The discriminative training of the MLP is performed by the feed-forward algorithm. The initialization of the MLPs weights is chosen randomly.