基于卷积神经网络的苗语孤立词语音识别

Hmong Language Speech Recognition for Isolated Words Based on Convolutional Neural Network

苗语作为西南少数民族的主要语种之一,其语音研究存在语言文字缺失、语音语料库不健全、地域差异等问题,因此苗语的语音识别难度较大。针对语音语料库不健全的问题,借助汉语拼音对苗语语音进行标注,构建不同地域的苗语语音语料库。基于拼音媒介传递的思想,设计苗语孤立词汇识别模型,该模型利用卷积神经网络对提取的语音特征进行训练和测试,解决了文字缺失情况下的苗语孤立词语音识别问题。为验证该模型的有效性和稳定性,以自建苗语语音语料库的数据作为实验数据集,分别在同地域、不同地域以及组合地域语音数据集上进行验证实验。实验结果表明,所提模型对苗语孤立词语音识别有效,同地域语音识别效果优于不同地域识别效果,在组合地域语音识别中表现亦良好。

Hmong language is one of the main languages of the inhabitants since of the Southwestern era,its phonetic research has problems such as lack of language and text,incomplete phonetic corpus,and regional differences. Therefore,the speech recognition of Hmong language challenges more. Aiming at the problem of imperfect speech corpus,this paper uses pinyin to label the Hmong language,and constructs a corpus of Hmong language voices in different regions;based on the idea of pinyin media transmission,a recognition model of Hmong language isolated words is designed,which uses convolutional neural networks. The network trains and tests the extracted speech features,and solves the problem of Hmong language isolated words speech recognition in the absence of text. In order to verify the effectiveness and stability of the proposed model,the self-built Hmong language speech corpus is used as the experimental data set,the experimental verification is carried out in the same region,different regions and combined regions respectively. The experimental results show that the proposed model is effective for the speech recognition of isolated words in Hmonglanguage,and the effect of speech recognition in the same area is better than that in different areas,The effect of speech experiment in combination area is good.