机构地区: 华南理工大学电子与信息学院
出 处: 《声学技术》 2007年第4期660-663,共4页
摘 要: KLT已经成功用于与文本无关的说话人辨认的特征提取,但是对于特征矢量分解,它需要巨大的计算负担。为了减轻计算负担,把KLT和重叠子帧合并起来用于噪声环境下的说话人辨认。基于重叠子帧的分离方法,文中提出了一种有效技术去建立特征矢量矩阵以便取得KLT的优点。在传统的MCE方法中,对于有K个说话人的系统而言,每一类别的分类错误都需要计算K-1类的判别函数,随着K的增加,使得计算量大量增加,文中提出改进的MCE模型以减少计算量,提高运算速度。实验结果显示:所提出的方法不仅减少了计算量,而且提高了系统辨认率。 KLT in feature extraction of text independent speaker identification needs huge computation load. To solve this problem, KLT is combined with overlap sub-frame for speaker identification in an additive noise environment. Based on a separation of overlap sub-frames, an effective technique is proposed to take the advantage of KLT. Experiments show that computation is reduced considerably. In the traditional MCE method, K-ldecision functions must be obtained for each classification error. With K increased, the computation load becomes prohibitive. A modified MCE model is described to alleviate the problem. Comparisons with GMM have been done, showing considerable improvement of the identification rate with KLT/MMCE. Especially when the hybrid number is up to 128, the system identification rate reaches 98.5%. This indicates effectiveness of the proposed method.