摘要
在佤语语音识别中,以孤立词作为识别单元时,未登录词对识别性能的影响很大。结合佤语语音特点,以音素作为识别单元,提出基于DNN-HMM声学模型的佤语语音识别方法。实验结果表明,与传统的GMM-HMM声学建模方法相比,基于DNN-HMM的声学模型表现出更加优越的识别性能,词错误率(WER)最优达29.24%。
In Wa language speech recognition,when isolated words are used as recognition units,unknown words have a great impact on recognition performance.Taking phonemes as recognition units,Wa language speech recognition method based on DNNHMM acoustic model is proposed.Experimental results show that compared with the traditional GMM-HMM acoustic modeling method,the proposed acoustic model shows better in recognition performance,and the word error rate(WER)is 29.24%.
作者
贾嘉敏
程振
潘文林
王欣
Jia Jiamin;Cheng Zhen;Pan Wenlin;Wang Xin(School of Mathematics and Computer Science,Yunnan Minzu University,Kunming,Yunnan 650500,China;School of Electrical Information Engineering,Yunnan Minzu University)
出处
《计算机时代》
2022年第8期61-64,68,共5页
Computer Era
基金
云南民族大学数学与计算机科学学院研究生科研项目(SJXY-2021-019)。
