In this paper, we proposed a hybrid MLP-HMM approach with tone recognition in order to improve the performance of Thai automatic speech recognition. The proposed approach consists of two main components: (1) a hybrid MLP-HMM as part of an acoustic model and (2) a tone feature extraction and classification using MLPs. The emission probabilities in the HMM framework are estimated by the posterior probabilities of neural network multilayer perceptrons in which the MFCC feature vectors are served into an input layer with fully-connected hidden layer. All nodes of output layer are represented by phonetic units. The MLPs are trained to estimate the emission probabilities for each state probability of the continuous density HMM framework. The Viterbi decoder is then used to find the single best state sequence using a dynamic programming algorithm. Compared to the baseline GMM system, the hybrid MLP-HMM significantly improved the performance of contextindependent networks. The recognition rates for consonants and vowels from the hybrid MLP-HMM model are higher than the baseline model, because the MLP model could help generate a recognition model which is more discriminative than the baseline HMM. In summary, the overall recognition rate was improved by using the proposed hybrid MLP-HMM approach.