基于深度神经网络的藏语语音关键词检索方法  

Akey word retrieval method for Tibetan speech based on deep neural network

作　　者：张恒 拉巴顿珠 官政先 肖鑫 Zhang Heng;Laba Dunzhu;Guan Zhengxian;Xiao Xin(School of Information Science and Technology,xizang informatization collaborative innovation center jointly built by the province and the Ministry,Xizang University,Lhasa 850000,China)

机构地区：[1]西藏大学信息科学技术学院西藏信息化省部共建协同创新中心,拉萨850000

出　　处：《西藏科技》2024年第6期73-80,共8页Xizang Science And Technology

基　　金：2022年西藏大学大学生创新性实验训练计划项目(2022XCX085)。

摘　　要：语音关键词识别作为人机语音交互的一项基础性研究课题,其目的是从连续的语音信号中提取特定的关键词,并实现对目标设备的唤醒以及其他相关功能。文章提出了一种基于DNN-HMM声学模型的藏语卫藏方言关键词检测方法。首先,通过切割、转换等方式对语音数据进行预处理;其次,使用MFCC从语音信号中提取出有效的特征作为模型的输入;再次,分别采用GMM-HMM和DNN-HMM模型对藏语声学特征进行建模。同时,为了提高模型的表现力和泛化能力,文章在模型中引入预训练和微调技术,对模型的结构进行了优化。实验结果表明,与传统基于GMM-HMM声学模型的识别结果相比,采用基于DNN-HMM声学模型的关键词检测方法能够更有效地检测出藏语语音关键词。As a basic research topic of human-computer voice interaction,speech keyword recognition aims to extract specific keywords from continuous speech signals and realize the wake-up of the target device and other related functions.This article proposes a keyword detection method for Tibetan-Amdo dialect based on the DNN-HMM acoustic model.Firstly,the speech data was preprocessed by cutting and converting.Secondly,MFCC was used to extract effective features from the speech signal as the input of the model.Finally,the GMM-HMM and DNN-HMM models were used to model the acoustic characteristics of Tibetan language respectively.At the same time,in order to improve the expressiveness and generalization ability of the model,this article introduces pre-training and fine-tuning techniques into the model,and optimizes the structure of the model.The experimental results show that compared with the traditional recognition results based on the GMM-HMM acoustic model,the keyword detection method based on the DNN-HMM acoustic model can detect Tibetan speech keywords more effectively.

关 键 词：声学模型 藏语 深度学习 关键词检测 语音识别 

分 类 号：TN912.3[电子电信—通信与信息系统] TP183[电子电信—信息与通信工程]