Abstract: 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.

Key words: Primi, deep learning, Kaldi speech recognition toolkit, speech recognition, robustness

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。

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

CLC Number: 

• TP18