Abstract: In this paper Parallel Hidden Markov Model (PHMM) made up of several par-allel Markov chains is proposed to fit in with speaker-independent speech recognition. The performance is improved because of the fusion of different models from classification based speech recognition. By sharing states of fused models, making classification automatically during training and getting result from all classifications, the amount of storage and operation can be decreased. The experiment for speaker-independent recognition of mandarin isolated digit shows that the PHMM improves the recognition performance and noise robustness.