用于文本无关的话者识别的超音段信息提取  

Extracting Supra-Segment Information for Text-Independent Speaker Verification

作　　者：姚志强[1] 吴礼福[1] 戴蓓蒨[1] 周曦[1] 

机构地区：[1]中国科技大学电子科学与技术系,合肥230027

出　　处：《数据采集与处理》2005年第4期376-380,共5页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(60272039)资助项目

摘　　要：提出一种可用于较少语音数据量的文本无关的超音段信息提取方法。通过对基音和能量的轨迹动态分段,提取超音段信息,并使用异方差线性区分分析(HLDA)进行参数优化,克服超音段信息提取对数据量大小的依赖,同时采用混合高斯-统一背景(GMM-U BM)模型结构,建立文本无关话者识别系统。在N IST′01数据库上的实验表明,该系统性能优于基于短时帧的音源信息参数系统,更重要的是不需要大数据量的支持,且与基于短时帧倒谱参数的话者识别系统融合后,系统识别性能明显改善,等误识率相对下降10%。A novel approach for extracting supra-segment information without large data is presented. Firstly, the pitch and energy contours are split into several segments. Then, parameters are extracted from each segment, and heteroscedastic linear discriminant analysis （HLDA） is used to parameter optimization. Finally, the Gaussian mixture model-universal background model （GMM-UBM） is used to establish the text-independent speaker verification system. Experiments on NISTr01 show that the system based on new supra-segmental parameters has a higher performance than the system based on the frame-level pitch and energy parameters. The recognition performance of the fustion system of the cepstrum based GMM is improved and the equal error rate is decreased by 10%.

关 键 词：话者确认 超音段信息 基音 文本无关 

分 类 号：TN912.34[电子电信—通信与信息系统]