Exploit posterior probability algorithm for pronunciation quality evaluation

In pronunciation quality evaluation, it is important to calculate the similarity between standard pronunciation and practical pronunciation. The posterior probability is the most stable and efficient for similarity measures. In order to improve the accuracy, this paper proposes a novel segmentation algorithm based on limited recognition networks and to calculate posterior probability in full probability space. In the meantime, for overcoming the shortage of poor discriminability of HMM, we optimize the full probability space based on linguistic knowledge. The experimental results show that the proposed algorithm is effective, the average scoring error rate (ASER) decreases 45.0% using the new phoneme segmentation algorithm and calculating posterior probability in optimal probability space.