干扰空间投影在本征音说话人自适应中的应用

APPLICATION OF NUISANCE SPACE PROJECTION IN EIGENVOICE SPEAKER ADAPTATION

本征音自适应是一种快速高效的自适应算法,它被广泛应用到说话人识别中,但由于同一个说话人的本征音自适应的说话人因子之间的信道特征和噪声存在差异,导致了算法的识别精度降低。针对这一问题,提出基于干扰空间投影的本征音说话人识别(EV-NSP)算法。将训练语音通过主成分分析(PCA)方法计算得到干扰投影矩阵;将投影矩阵应用到生成本征音矢量算法中;利用最大似然估计算法自适应地得到说话人因子的估计值。实验结果表明,EV-NSP算法相对于传统的本征音自适应算法识别性能有了较大的提高。

However, Eigenvoice adaption is a fast and efficient adaptive algorithm which is widely used the speaker factor of eigenvoice adaption from the same speaker has different channel c in speaker recognition. haracteristies and noise feature, thus reduced the recognition accuracy. To solve this problem, we propose an eigenvoice speaker recognition algorithm based on nuisance space projection （EV-NSP）. We calculated the nuisance space projection matrix by principal component analyzing training data, and then run the eigenvoice vector algorithm by using projection matrix. The maximum likelihood estimation algorithm was used to adaptively estimate the speaker factor. Experimental results indicate that EV-NSP algorithm have better performance compared with traditional adaptive eigenvoice algorithm.