基于有效KLT和重叠子帧的说话人辨认(英文)

Based on Effective KLT and Overlap Sub-frame for Speaker Identification

把KLT和重叠子帧用于在噪声环境下的说话人辨认。基于重叠子帧的分离方法,提出了一种有效技术去建立特征矢量矩阵和取得KLT技术的优点的有效性。做了几个实验。和GMM相比较,采用KLT/MMCE的辨认率得到了明显的提高。特别是当混合数达到128时,辨认率达到了98.5%。因此,实验结果现实所提出的方法确实能减少计算量和提高系统的辨认率。

Karhunen Looeve transform （KLT） and overlap sub-frame are used to the effective speaker identification in the additive noise environment. Based on the separate method, an effective techniqued is proposed to establish feature matrix and acquire the validity of KLT technique. At the recognition stage, a kind of modified MCE model is proposed to decrease computation capacity and further to enhance computation velocity. Several comparing experiments have been done. Compared with GMM, the identification rate adopted KLT/MMCE improves obviously. Especially when the hybrid number is up to 128, the system identification rate achieves 98.5%. Therefore, the experiment result is showed that the proposed method can indeed decrease the computation capacity and improve the identification rate of the system.