An acoustic model for a real-time continuous phoneme recognition system must exhibit the following desirable feature: an ability to minimize the recognition performance degradation while solving the model complexity problem to confine the delay to a minimum in recognition process. To cope with the challenges, we introduce the state-dependent Phonetic Tied-Mixture (PTM) model with Head-Body- Tail (HBT) structured HMM as an acoustic model optimization. The proposed acoustic modeling method shows a significant improvement in recognition performance and becomes a solution to the sparse training data problem and the model complexity problem. Moreover, defining the exceptional Gaussian mixtures in tying process achieves a drastic reduction in phoneme error rate compared to traditional state-dependent PTM method. In this paper, we describe the new acoustic model optimization procedure and show the outstanding performance evaluation results for real-time continuous phoneme recognition system.