声纹识别中的语言属性映射

短时声纹识别中语言属性不匹配问题降低了话者识别系统性能，语言属性失配在短时话者识别中的影响成为当前研究热点。从消除特征参数中的语言属性信息出发，提出了一种语言属性映射方法。采用均值超向量来构建训练样本，应用高斯混合模型对每一句话不等长的特征进行建模，引入加权矩阵得到语言属性空间，通过映射方法在统计参数超矢量空间中消除语言属性的影响。使用当前两套经典声纹识别系统作为基线系统做对比试验，结果表明：相对基线系统在等错误率上，男声测试部分降低了23.07%，女声测试部分降低了20.39%，从而验证了正确性和有效性。

Speaker verification based on linguistic attribute projection

Linguistic attribute mismatch reduces the performance of speaker recognition system in short-time voiceprint recognition, so linguistic attribute mismatch has become the hot spot in research. A linguistic attribute projection approach was proposed to eliminate the linguistic attribute information in acoustic feature parameters. We built train samples with mean supervector, and used gauss mixture model for unequal length feature of every word. Moreover, the introduction of weighted matrix gets linguistic attribute space, and eliminates the effects of linguistic attribute in statistical parameters supervector space through projection method. Experimental results demonstrate that the equal error rate of the proposed system can be reduced by 23.07% and 20.39% against the baseline system in male and female trials part respectively.