深度神经网络技术在汉语语音识别声学建模中的优化策略

发布时间：2018-05-09 13:44

本文选题：深层神经网络 + 语音识别　；参考：《重庆邮电大学学报(自然科学版)》2014年03期

【摘要】：将深度神经网络作为声学模型引入面向汉语电话自然口语交谈语音识别系统。针对自然口语中识别字错误率较高的问题,从语音的声学特征类型选择、模型训练时元参数调节以及改善模型泛化能力等方面出发,对基于深度神经网络的声学模型建模技术进行了一系列的优化。针对训练样本中状态先验概率分布稀疏的情况,提出了一种状态先验概率平滑算法,在一定程度上缓解了这种数据稀疏问题,经平滑后,字错误率下降超过1%。在所采用的3个电话自然口语交谈测试集上,相对于优化前的深度神经网络模型,经过优化后的模型取得了性能的一致提升,字错误率平均相对降低15%。实验结果表明,所采用优化策略可以有效地改善深度神经网络声学模型性能。
[Abstract]:The depth neural network is used as the acoustic model in the speech recognition system for natural spoken conversation of Chinese telephone. Aiming at the problem of high recognition error rate in natural spoken language, this paper starts with the selection of acoustic feature types of speech, the adjustment of model training time element parameters, and the improvement of model generalization ability, etc. The acoustic modeling technology based on depth neural network is optimized. A state priori probability smoothing algorithm is proposed for sparse state prior probability distribution in training samples. To some extent, the problem of data sparsity is alleviated. After smoothing, the word error rate decreases by more than 1 bit. Compared with the depth neural network model before the optimization, the performance of the optimized model is consistently improved, and the word error rate is reduced by 15%. Experimental results show that the proposed optimization strategy can effectively improve the performance of depth neural network acoustic model.
【作者单位】：中国科学院语言声学与内容理解重点实验室;
【基金】：国家自然科学基金(10925419,90920302,61072124,11074275,11161140319,91120001,61271426) 中国科学院战略性先导科技专项(XDA06030100,XDA06030500) 国家“863”计划(2012AA012503) 中科院重点部署项目(KGZD-EW-103-2)~~
【分类号】：TN912.34;TP183

【参考文献】