跨语言声学模型在维吾尔语语音识别中的应用  

Crosslingual acoustic modeling in Uyghur speech recognition

作　　者：努尔麦麦提·尤鲁瓦斯[1] 刘俊华 吾守尔·斯拉木[1] 热依曼·吐尔逊[1] 达吾勒·阿布都哈依尔[1] NURMEMET Yolwas;LIU Junhua;WUSHOUR Silamu;REYIMAN Tursun;DAWEL Abilhayer(College of Information Science and Engineering Xinjiang University, Urumqi 830046, China;iFLYTEK Co. , Ltd. , Hefei 230088, China)

机构地区：[1]新疆大学信息科学与工程学院,乌鲁木齐830046 [2]科大讯飞股份有限公司,合肥230088

出　　处：《清华大学学报（自然科学版）》2018年第4期342-346,共5页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金项目(61363063,U1603262,61462084);新疆维吾尔自治区重点实验室项目(2015KL013)

摘　　要：对维吾尔语而言,由于数据采集和标注存在各种困难,用于训练声学模型的语音数据不够充分。为此,该文研究了基于长短期记忆网络的跨语言声学模型建模方法,利用汉语庞大的训练数据训练深度神经网络声学模型,然后将网络的输出层权重去掉,用随机化的方式产生与维吾尔语输出层对应的权重值,采用反向传播的方式,利用维吾尔语语音数据更新所有权重来训练维吾尔语声学模型。实验结果表明：该方法使维吾尔语转写和听写识别错误率分别比基线系统相对降低了20%和30%。该方法利用汉语大数据来训练神经网络的隐藏层,使维吾尔语声学模型能在一个较好的初始权重网络上进行训练,增强了网络的鲁棒性。The Uyghur language has a little speech data for training acoustic models due to various data acquisition and annotation difficulties.This paper describes a modeling method for crosslingual acoustic models based on long short-term memory models.Mass Chinese language training data is used to train a deep neural network acoustic model.The network output layer weights are then randomly modified to create the output layer for the Uyghur language.A Uyghur language acoustic model is then trained using Uyghur language speech data to update all the weights.Tests show that this method reduces the word error rates of the Uyghur language transcription and dictation recognition by 20% and 30% than the baseline system.Thus,this method improves the Uyghur language acoustic model with better initial weights from the Chinese language data to train hidden layers in the neural network,and enhances the network robustness.

关 键 词：声学模型 维吾尔语 跨语言 长短期记忆 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]