基于神经网络的普通话学习者声调自动检测研究  

Research on Automatic Tone Detection for Mandarin Learners Based on Neural Networks

作　　者：古力努尔·艾尔肯 艾斯卡尔·艾木都拉[2] GULNUR Arkin;ASKAR Hamdulla(Institute of Public Administration Xinjiang University of Finance and Economics,Urumqi,830012,China;School of Intelligent Science and Technology,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆财经大学公共管理学院,乌鲁木齐830012 [2]新疆大学智能科学与技术学院,乌鲁木齐830046

出　　处：《自动化与仪器仪表》2024年第12期153-158,共6页Automation & Instrumentation

基　　金：国家自然科学基金项目资助(62307030)。

摘　　要：声调是汉语普通话中最重要的信息,随着汉语语音技术的进一步发展,声调的自动检测已经成为汉语语音技术发展的一个主要方向。本研究以标准和非标准的汉语普通话发音的大规模中介语料库为基础数据,采用基于深度神经网络(Deep Neural Network,DNN)不同特征的训练方法,即基频F0特征、39维MFCC(Mel-Frequency Cepstral Coefficients,MFCC)特征和由基频F0和39维MFCC融合的参数特征等三类不同的方法,进一步对学习者的语音声调进行自动检测。实验结果表明,由基频F0和39维MFCC特征融合方法对网络模型的训练效果最优。另外,为进行与DNN模型的对比,运用相同的特征参数进行了高斯混合模型(Mixture Gaussian Model,GMM)-隐马尔科夫模型(Hidden Markov Model,HMM)模型的训练,并获得了良好的对比实验结果。最后,为使实验结果更加科学、准确,还通过声调感知实验来考察学习者对汉语声调的感知情况。该研究提供了一种针对普通话声调学习的客观而有效的评估方法,有助于提升学习者国家通用语言学习效率。Tone is the most important information in Mandarin.With the further development of Chinese phonetic technology,automatic tone detection has become a main direction of the development of Chinese phonetic technology.This study is based on a large-scale intermediary corpus of standard and nonstandard Mandarin pronunciation,the training methods based on different features of deep neural network(DNN),there are three different methods:F0 feature of fundamental frequency,39 dimensional MFCC feature and parameter feature fused by fundamental frequency F0 and 39 dimensional MFCC,are used to detect the learners'tone automatically.Therefore,the characteristics composed of fundamental frequency F0 and MFCC parameters are the best for the training effect of the network model.In addition,in order to compare with the DNN model,the GMM-HMM model was trained with the same characteristic parameters,and good experimental results were obtained.Finally,use tone perception experiments to investigate learners’perception of Mandarin tones.This research provides an objective and effective evaluation method for Mandarin tone learning,which helps improve the efficiency of learners Mandarin.

关 键 词：深度神经网络 普通话学习者 声调自动检测 计算机辅助语言系统 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]