Summary: By combining the phone posterior and phone accuracy, a data selection method based on variable weighting is proposed to improve the discriminative training performance of the acoustic models for continuous speech recognition. Firstly, the word lattice is reduced by using a posterior-based Beam pruning method, and for each hypothesis word a weight is derived from the word error rates of the path containing that word with the posterior. Then, each pair of confusing phones is variably weighted according to a phone confusion matrix, and the modified phone accuracy is calculated by applying those weights. Finally, the distribution of the expected phone accuracies is estimated and all competing arcs are soft weighted using Gaussian functions. Experimental results show that compared with the minimum phone error criterion, the variable weighting method not only improves the recognition rate by 0.61%, but also reduces the required training time.