一种适于说话人辨认的自适应频率尺度变换  

Adaptive Frequency Transform for Speaker Identification

作　　者：李燕萍[1,2] 唐振民[1] 丁辉[1,3] 张燕[1,4] 

机构地区：[1]南京理工大学计算机科学与技术学院,江苏南京210094 [2]南京邮电大学通信与信息工程学院,江苏南京210003 [3]嘉兴学院数学与信息工程学院,浙江嘉兴314001 [4]金陵科技学院信息技术学院,江苏南京210006

出　　处：《南京理工大学学报》2010年第2期182-186,共5页Journal of Nanjing University of Science and Technology

基　　金：浙江省自然科学基金(Y1090649);浙江省教育厅科研资助项目(Y200805349);南京邮电大学引进人才科研启动基金(NY209004)

摘　　要：该文提出了一种适于说话人辨认的自适应频率尺度变换,基于说话人信息在不同频带中的非均匀分布性质,通过F比衡量不同频率子带对说话人信息的贡献大小,设计自适应频率滤波器,提高贡献大的频带的频率分辨率,降低贡献小的频带的频率分辨率,提取鉴别性特征DFCC。干净语音环境下,不同测试文件的实验表明,该文提出的DFCC特征的识别率比传统MFCC特征平均提高了1.45%,表明特征的稳定性好,对语音内容不存在依赖性;在不同信噪比的噪声环境下,识别率平均提高了6.37%,表明DFCC特征能够充分利用语音频带中包含的说话人信息,具有良好的抗噪性能。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.

关 键 词：说话人辩认 自适应频率尺度变换 鉴别性特征 非均匀子带 

分 类 号：TN912.34[电子电信—通信与信息系统]