In this paper, Neural Predictors (NP) are introduced to Hidden Markov Models (HMM) to form a new model HMM / NP for speech recognition. In HMM/NP, each NP is a Multilayer Perceptron (MLP) used as a separate nonlinear predictor, and corresponds to a state in the model. Training and recognition algorithms are given based on Baum-Welch and Back-Propagation (BP) algorithms.Speaker-dependent Mandarin digit recognition experiments were carried out. The performance of forward prediction HMM / NP model and forward-backward prediction HMM / NP model was comparied and a recognition accurcy of 96.2% and 98.7% was obtained respectively. The result indicates that it is effective to use HMM/NP for Mandarin speech recognition.