一种区分性互补系统构造与融合的语音识别方法

发布时间：2018-03-25 16:23

本文选题：模型层　切入点：语音识别　出处：《声学学报》2016年01期

【摘要】：在区分性训练的框架下,提出了一种基于混淆信息加权的互补系统构造方法。首先通过统计音素对的混淆信息,利用混淆信息给音素对加以不同的惩罚权重,分别以基线系统中的3个最优识别结果作为参考,计算混淆信息加权后的音素准确率,同时以正确的标注为参考计算标准的音素准确率。然后通过同时最大化混淆信息加权后的音素准确率和最小化标准音素准确率,构建模型层互补系统,并进一步通过结合RDLT(region-dependent linear transform)特征变换过程构造特征层的互补系统。实验结果表明,与互补最小音素错误准则相比,融合模型层互补系统后识别率提高了0.76%,同时融合特征层和模型层的互补系统识别率提高了1.35%。本方法可以增大互补系统间的差异性,提高系统融合后的识别性能。
[Abstract]:In the framework of discriminative training, a complementary system construction method based on weighted confounding information is proposed. Firstly, by using the confounding information of phoneme pairs, the confounding information is used to give different penalty weights to phoneme pairs. Based on the three optimal recognition results in the baseline system, the phoneme accuracy of weighted confounding information is calculated. At the same time, the correct tagging is used as the reference to calculate the phoneme accuracy. Then, by maximizing the weighted phoneme accuracy of confused information and minimizing the standard phoneme accuracy, a model level complementary system is constructed. Furthermore, the complementary system of the feature layer is constructed by combining the RDLT(region-dependent linear transform process. The experimental results show that, compared with the complementary minimum phoneme error criterion, The recognition rate of complementary system at model level is improved by 0.76 and the recognition rate of complementary system by combining feature layer and model layer is improved by 1.35. This method can increase the difference between complementary systems and improve the recognition performance after system fusion.
【作者单位】：解放军信息工程大学信息系统工程学院;
【基金】：国家自然科学基金(61175017) 国家高技术研究发展计划(863)(2012AA011603)资助
【分类号】：TN912.34

【参考文献】