基于Kaldi的普米语语音识别  

Primi Speech Recognition Based on Kaldi

作　　者：胡文君[1] 傅美君 潘文林[1] 

机构地区：[1]云南民族大学数学与计算机科学学院,昆明650500

出　　处：《计算机工程》2018年第1期199-205,共7页Computer Engineering

基　　金：国家科技支撑计划项目(2013BAJ07B02-1);云南省教育厅科学研究基金(2016YJS078);云南省高校物联网应用技术重点实验室开放研究课题(2015IOT02)

摘　　要：为提高普米语语音识别系统的性能,引入深度学习模型进行普米语语音识别,该模型是一个高容量复杂的网络模型。以Kaldi语音识别工具包为实验平台,分别训练5种不同的声学模型,且这5种模型中包含一个有4隐层的深度神经网络模型。比较不同声学模型得到的语音识别率发现,G-DNN模型比Monophone模型的语音识别率平均提升49.8%。实验结果表明,当增加训练集的普米语语音语料量时,基于深度学习的普米语语音识别率会提升,而基于深度学习的普米语语音识别系统的鲁棒性比其余4个声学模型的普米语语音识别系统的鲁棒性更强。In order to improve the performance of Primi speech recognition system, the deep learning model is introduced into Primi speech recognition. The deep learning model is a large capacity and complex network model. Kaldi speech recognition toolkit is used as an experimental platform and five different acoustic models are respectively trained which contain a deep neural network model with four hidden layers. By comparing the speech recognition rates obtained by different acoustic models,it is found that the G-DNN model improves the accuracy of speech recognition by 49.8% than the Monophone model. Experimental results show that the Primi speech recognition rate based on the deep learning model can be improved, when the number of Primi speech corpus in the training set is increased. And the robustness of the Primi speech recognition system based on deep learning is stronger than the other four acoustic models.

关 键 词：普米语 深度学习 Kaldi语音识别工具包 语音识别 鲁棒性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]