|Table of Contents|

Adaptive Frequency Transform for Speaker Identification


Research Field:
Publishing date:


Adaptive Frequency Transform for Speaker Identification
LI Yan-ping14TANG Zhen-min1DING Hui12ZHANG Yan13
1.School of Computer Science and Technology,NUST,Nanjing 210094,China;2.School of Mathematics and Information Engineering,Jiaxing University,Jiaxing 314001,China;3.School of Information Technology,Jinling Institute of Technology,Nanjing 210006,China;4.College of Telecommunication and Information Engineering,Nanjing Univesity of Posts andTelecommunications,Nanjing 210003,China
speaker identification adaptive frequency transform discriminative feature non-uniform sub-bands
A novel method for speaker identification based on adaptive frequency transform is proposed here.According to the fact that the speaker information is non-uniformly distributed in frequency bands,the discrimination power between frequency components and individual characteristics is examined and the speaker information is quantified based on Fisher’s F-ration.A new adaptive frequency filter is designed,which can improve the frequency resolution in high contribution frequency domain,reduce the frequency resolution in low contribution frequency domain,and extract the discriminative feature DFCC(Discriminative frequency cepstral coefficient).In a clean environment,the results from the experiments on different testing materials indicate that the recognition rates based on DFCC increases by 1.45% on average than on traditional MFCC(Mel frequency cepstral coefficient),which confirms that the proposed feature is stable and independent of spoken contents.Furthermore,in the noise environment of different SNR levels,the experiment results demonstrate that the recognition rate increases by 6.37% on average,which confirms the effectiveness of discrimination and robustness of DFCC.


[1]Campbell J P. Speaker recognition: a tutoria l[ J]. Proceed ings o f the IEEE, 1997, 85( 9): 1437- 1462.
[2] H ayakawa S, Itakura F. Tex t-dependent speake r recogn ition using the inform a tion in the higher frequency band[ A]. Proceed ings o f the Conference on Acoustic, Speech and S igna l Pro cessing [ C ]. Adela ide, SA, Australia, IEEE, 1994: 19- 22.
[3] M iyajmi a C, Watanab leH, Tokuda K, et a.l A new approach to designing a feature extractor in speaker identif-i cation based on discrmi inative feature ex traction [ J]. Speech Commun ication, 2001, 35( 3): 203- 218.
[4] Lu Xugang, Dang Jianwu. An investigation o f dependenc ies between frequency components and speaker character istics for tex-t independent speaker identification [ J]. Speech Commun ication, 2008, 50: 312- 322.
[5] 俞一彪, 袁冬梅, 薛峰. 一种适于说话人识别的非 线性频率尺度变换[ J]. 声学学报, 2008, 33 ( 5): 450- 455.
[6] 赵力. 语音信号处理[M ]. 北京: 机械工业出版 社, 2008.
[7] Reyno lds D A, Rose R C. Robust tex t- independent speaker identifica tion using Gaussian m ix ture speaker m ode ls[ J]. IEEE Transac tions on Speech and Audio Processing, 1995, 3( 1): 72- 83.
[8] Dang J, H onda K. Acoustic character istics o f the p ir-i fo rm fossa in models and hum ans[ J] . Acoustica l So c-i ety o f Am er ica, 1997, 101: 456- 465.
[9] K itam ura T, H ondaK, Takem o to H. Ind iv idua l variation of the hypopha ryngea l cav ities and its acoustic e ffects [ J] . A coustical Soc ie ty of Am er ica, 2005, 26( 1): 16- 26.
[10] ChanW N, Zheng N, Lee T. D iscrim ination power of voca l source and vocal tract related fea tures for speaker segm en tation[ J]. IEEE Transactions on Audio, Speech and Language Process ing, 2007, 15( 6): 1884- 1892.
[11] Varga A, Steeneken H JM, Tom linson M, et a.l The NOISEX-92 study on the effect o f add ictive noise on automa tic speech recogn ition[ R ]. Ma lve rn, UK: Speech Research Un it, Defense Research Agency, 1992. 186


Last Update: 2010-04-30