采用长短时记忆网络的低资源语音识别方法

发布时间：2018-03-15 20:09

本文选题：语音识别　切入点：低资源　出处：《西安交通大学学报》2017年10期 　论文类型：期刊论文

【摘要】：针对低资源环境下由于标注训练数据不足、造成语音识别系统识别率急剧下降的问题,提出一种采用长短时记忆网络的低资源语音识别(LSTM-LRASR)方法。该方法采用长短时记忆网络构建声学模型,从特征提取、数据扩展及模型优化3个方面提高低资源语音识别性能。在特征提取方面,提取语言无关的高层稳健特征参数,降低声学模型对训练数据的依赖;在数据扩展方面,对已有标注数据进行语速扰动,对无标注数据进行自动识别,从而自动获取更多标注数据;在模型优化方面,通过序贯区分性训练技术提高模型对易混淆音素的区分能力,利用最小风险贝叶斯解码对多个系统进行融合,进一步提高识别性能。对OpenKWS16评测数据的实验结果表明,采用LSTMLRASR方法搭建的低资源语音识别系统的词错率相对基线系统下降了29.9%,所有查询词的查询项权重代价提升了60.3%。
[Abstract]:Aiming at the problem that the recognition rate of speech recognition system drops sharply due to the shortage of training data in low resource environment, In this paper, a LSTM-LRASR method based on LSTM-LRASR is proposed, which uses LSTM-LRASR to construct acoustic model and extract features from LSTM-LRASR. Data expansion and model optimization improve the performance of low resource speech recognition. In feature extraction, we extract high-level robust feature parameters independent of language to reduce the dependence of acoustic model on training data. In the aspect of model optimization, the ability of model to distinguish confusing phoneme is improved by sequential discriminant training technology. The minimum risk Bayesian decoding is used to fuse multiple systems to further improve the recognition performance. The experimental results of OpenKWS16 data evaluation show that, The word-error rate of the low resource speech recognition system based on LSTMLRASR method is 29.9m lower than that of the baseline system, and the cost of the query item weight of all the query words is increased by 60.3%.
【作者单位】：解放军信息工程大学信息系统工程学院;中国科学技术大学信息科学技术学院;
【基金】：国家自然科学基金资助项目(61673395,61403415,61302107) 河南省自然科学基金资助项目(162300410331)
【分类号】：TN912.34

【相似文献】