论文部分内容阅读
在说话人确认的任务中,为了解决信道失配问题,提高系统性能,引入了联合因子分析和稀疏表示算法。首先利用联合因子分析算法去除信道干扰,得到与信道无关的说话人因子,然后在稀疏表示算法中利用说话人因子构建过完备字典,求解稀疏最优化问题计算说话人得分。由于此方法有机结合了联合因子分析算法的信道鲁棒性和稀疏表示的鉴别性,使用此算法构建的系统在NIST SRE 2008电话训练、电话测试数据集上性能表现良好,相对于联合因子分析-支持向量机系统在性能上有竞争性,在原理上有互异性,系统融合更带来了最小检测代价指标上4.91%的性能提升。实验表明使用联合因子分析与稀疏表示进行说话人确认是可行的。
In the task of speaker recognition, in order to solve the channel mismatch problem and improve the system performance, the joint factor analysis and sparse representation algorithm are introduced. Firstly, the channel interference is removed by the joint factor analysis algorithm to get the channel independent speaker factor. Then, the sparse representation algorithm is used to construct the perfect dictionary by using the speaker factor, and the sparse optimization problem is solved to calculate the speaker score. Because this method organically combines the channel robustness of the joint factor analysis algorithm and the discriminability of the sparse representation, the system constructed by using this algorithm performs well on the NIST SRE 2008 telephone training and telephone test dataset. Compared with the joint factor analysis - Support vector machine system in the performance of competitive, in principle, there are differences, the system integration brings the minimum detection cost 4.91% performance index. Experiments show that the use of joint factor analysis and sparse representation for speaker verification is feasible.