论文部分内容阅读
提出一种利用自适应语料和训练语料对模型状态结构调整的算法。该算法在易混淆的状态间参数共享,提高了模型对样本的后验概率和对自适应语料的利用率,并间接地调整了系统决策树的结构。识别实验结果表明,在不同数量的自适应语句下,调整后的系统识别率比基线系统获得了一致的提高,结合使用MLLR说话人自适应,调整的系统识别率平均提高了 15．60％,有效地减少了测试语料与训练语料决策树结构不匹配造成的系统识别率降低。
This paper proposes an algorithm to adjust the state structure of the model by using adaptive corpus and training corpus. The algorithm shared parameters in a confusing state, which improved the model posterior probability and adaptive corpus utilization, and indirectly adjusted the structure of the system decision tree. The experimental results show that under different numbers of adaptive sentences, the adjusted system recognition rate has been improved consistently compared with the baseline system. Combined with MLLR speaker adaptation, the adjusted system recognition rate has been increased by 15.60% on average, Effectively reducing the system recognition rate caused by the mismatch between the test corpus and the training corpus decision tree structure.