文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Speaker Verification with Adaptive Spectral
侵权投诉

Speaker Verification with Adaptive Spectral

Abstract. Spectral subband centroids (SSC) have been used as an additional feature to cepstral coefficients in speech and speaker recognition. SSCs are computed as the centroid frequencies of subbands and they capture the dominant frequencies of the short-

SpeakerVeri cationwithAdaptiveSpectral

SubbandCentroids

TomiKinnunen1,BingjunZhang2,JiaZhu2,andYeWang2

SpeechandDialogueProcessingLab

InstitutionforInfocommResearch(I2R)

21HengMuiKengTerrace,Singapore119613

ktomi@i2r.a-star.edu.sg

2DepartmentofComputerScience

SchoolofComputing,NationalUniversityofSingapore(NUS)

3ScienceDrive2,Singapore117543

{bingjun,zhujia,wangye}@comp.nus.edu.sg1

Abstract.Spectralsubbandcentroids(SSC)havebeenusedasanaddi-tionalfeaturetocepstralcoe cientsinspeechandspeakerrecognition.SSCsarecomputedasthecentroidfrequenciesofsubbandsandtheycapturethedominantfrequenciesoftheshort-termspectrum.InthebaselineSSCmethod,thesubband ltersarepre-speci ed.Toallowbet-teradaptationtoformantmovementsandotherdynamicphenomena,weproposetoadaptthesubband lterboundariesonaframe-by-framebasisusingagloballyoptimalscalarquantizationscheme.Themethodhasonlyonecontrolparameter,thenumberofsubbands.Speakerveri- cationresultsontheNIST2001taskindicatethattheselectionoftheparameterisnotcriticalandthatthemethoddoesnotrequireadditionalfeaturenormalization.

1Introduction

Theso-calledmel-frequencycepstralcoe cients[1](MFCC)haveproventobee cientfeaturesetforspeakerrecognition.Aknownproblemofcepstralfeatures,however,isnoisesensitivity.Forinstance,convolutivenoiseshiftsthemeanvalueofthecepstraldistributionwhereasadditivenoisetendstomodifythevariances

[2].Tocompensateforthefeaturemismatchbetweentrainingandveri cationutterances,normalizationsinfeature,modelandscoredomainsarecommonlyused[3].

Spectralsubbandcentroids[4;5;6;7](SSC)areanalternativetocepstralcoe cients.SSCsarecomputedasthecentroidfrequenciesofsubbandspectraandtheygivethelocationsofthelocalmaximaofthepowerspectrum.SSCshavebeenusedforspeechrecognition[4;5],speakerrecognition[7]andaudio ngerprinting[6].RecognitionaccuracyofSSCsislowerinnoise-freecondi-tionscomparedwithMFCCs.However,SSCscanoutperformMFCCsinnoisyconditionsandtheycanbecombinedwithMFCCstoprovidecomplementaryinformation.

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email