论文部分内容阅读
提出了基于发音特征单个维度的置信度算法,并基于此对发音特征的各个维度展开分析。分析不仅验证了融合的必要性,同时也展示了发音特征各维度之间以及和隐马尔可夫模型之间的大量冗余。为了去除冗余,提出了用子集选择的方法进行优化。对比所有都用的情况,基于发音特征紧凑子集的语音识别置信度估计,在等错率上取得了12.7%的相对下降。把经过优化后的基于发音特征的语音识别置信度估计和基于隐马尔可夫模型的语音识别置信度进行融合,在保持集内识别率不损失的前提下,显著提高了语法外输入测试的拒识性能:在相同参数下,在开发集和测试集上分别取得了34%和35.3%的显著改善。
A confidence algorithm based on a single dimension of phonetic features is proposed, and based on this, the analysis of each dimension of phonetic features is carried out. The analysis not only verifies the necessity of fusion, but also shows a great deal of redundancy between dimensions of pronunciation features and hidden Markov models. In order to remove the redundancy, a subset selection method is proposed to optimize. In contrast to all cases, the confidence of speech recognition based on the compact subset of phonetic features has achieved a relative decline of 12.7% at equal error rates. Combining the optimized confidence estimation based on pronunciation feature and the confidence recognition based on hidden Markov model, the recognition rate of grammar input is significantly improved without losing the recognition rate in the set Performance: With the same parameters, significant improvements of 34% and 35.3% were achieved on the development set and test set, respectively.