基于MTL-DNN系统融合的混合语言模型语音识别方法

发布时间：2018-04-08 22:11

本文选题：集外词　切入点：混合模型　出处：《数据采集与处理》2017年05期

【摘要】：基于混合语言模型的语音识别系统虽然具有可以识别集外词的优点,但是集外词识别准确率远低于集内词。为了进一步提升混合语音识别系统的识别性能,本文提出了一种基于互补声学模型的多系统融合方法。首先,通过采用不同的声学建模单元,构建了两套基于隐马尔科夫模型和深层神经网络(Hidden Markov model and deep neural network,HMM-DNN)的混合语音识别系统;然后,针对这两种识别任务之间的关联性,采用多任务学习(Multi-task learning DNN,MTL-DNN)思想,实现DNN网络输入层和隐含层的共享,并通过联合训练提高建模精度。最后,采用ROVER(Recognizer output voting error reduction)方法对两套系统的输出结果进行融合。实验结果表明,相比于单任务学习DNN(Single-task learning DNN,STL-DNN)建模方式,MTL-DNN可以获得更好的识别性能;将两个系统的输出进行融合,能够进一步降低词错误率。
[Abstract]:Although the speech recognition system based on the hybrid language model has the advantage of recognizing extrinsic words, the accuracy of extrinsic words recognition is much lower than that of intra-set words.In order to improve the recognition performance of hybrid speech recognition system, a multi-system fusion method based on complementary acoustic model is proposed in this paper.Firstly, two hybrid speech recognition systems based on Hidden Markov model and deep neural Network (HMM-DNN) are constructed by using different acoustic modeling units.Multi-task learning learning MTL-DNN is adopted to realize the sharing of input layer and hidden layer in DNN network, and the modeling accuracy is improved by joint training.Finally, the output results of the two systems are fused by ROVER(Recognizer output voting error reduction method.The experimental results show that the MTL-DNN can obtain better recognition performance than the single-task learning DNN(Single-task learning DNNN (STL-DNNN) modeling method, and can further reduce the word error rate by merging the output of the two systems.
【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(61175017,61403415,61302107)资助项目
【分类号】：TN912.34

【相似文献】