本文選題：發(fā)音特征　切入點(diǎn)：發(fā)音偏誤趨勢(shì)　出處：《北京大學(xué)學(xué)報(bào)(自然科學(xué)版)》2017年02期 　論文類型：期刊論文

【摘要】：為了提升計(jì)算機(jī)輔助發(fā)音訓(xùn)練(CAPT)系統(tǒng)中發(fā)音偏誤趨勢(shì)(PET)的檢測(cè)效果,確保反饋信息的準(zhǔn)確性與有效性,提出一種基于對(duì)數(shù)似然比的發(fā)音特征方法。該方法將多個(gè)基于深度神經(jīng)網(wǎng)絡(luò)的發(fā)音特征提取器用于生成幀級(jí)別的對(duì)數(shù)似然比,然后將對(duì)數(shù)似然比組成的發(fā)音特征用于PET的檢測(cè),為學(xué)習(xí)者提供發(fā)音位置和發(fā)音方法的正音信息。實(shí)驗(yàn)結(jié)果表明,發(fā)音特征對(duì)PET的檢測(cè)效果優(yōu)于常用聲學(xué)特征(MFCC,PLP和f Bank),當(dāng)發(fā)音特征與MFCC特征相結(jié)合時(shí),可以進(jìn)一步提升性能,達(dá)到錯(cuò)誤接受率為5.0%,錯(cuò)誤拒絕率為30.8%,診斷正確率為89.8%的檢測(cè)效果。
[Abstract]:In order to improve the detection effect of pronunciation bias trend in CAPTT system, and ensure the accuracy and effectiveness of feedback information, A method of pronunciation feature based on logarithmic likelihood ratio (LLR) is proposed, in which several speech feature extractors based on depth neural network are used to generate logarithmic likelihood ratio at frame level, and then the pronunciation feature composed of logarithmic likelihood ratio is used for PET detection. The experimental results show that pronunciation features are more effective than common acoustic features, such as PET and f BankP, and can further improve the performance when the pronunciation features are combined with MFCC features. The error acceptance rate is 5.0%, the error rejection rate is 30.8%, and the diagnostic accuracy rate is 89.8%.
【作者單位】： 北京語(yǔ)言大學(xué)信息科學(xué)學(xué)院;
【基金】：北京語(yǔ)言大學(xué)梧桐創(chuàng)新平臺(tái)項(xiàng)目(16PT05)和北京語(yǔ)言大學(xué)研究生創(chuàng)新基金項(xiàng)目(16YCX160)資助
【分類號(hào)】：H01;TP391.7

【相似文獻(xiàn)】

相關(guān)會(huì)議論文 前1條

1 盧東斌;李光球;江虹;;基于比特對(duì)數(shù)似然比GSC的MQAM性能分析[A];浙江省電子學(xué)會(huì)2006年學(xué)術(shù)年會(huì)論文集[C];2006年

，

本文編號(hào)：1576528