论文部分内容阅读
提出了一种文本无关说话人识别的全特征矢量集模型及互信息评估方法,该模型通过对一组说话人语音数据在特征空间进行聚类而形成,全面地反映了说话人语音的个性特征。对于说话人语音的似然度计算与判决,则提出了一种互信息评估方法,该算法综合分析距离空间和信息空间的似然度,并运用最大互信息判决准则进行识别判决。实验分析了线性预测倒谱系数(LPCC)和Mel频率倒谱系数(MFCC)两种情况下应用全特征矢量集模型和互信息评估算法的说话人识别性能,并与高斯混合模型进行了比较。结果表明:全特征矢量集模型和互信息评估算法能够充分反映说话人语音特征,并能够有效评估说话人语音特征相似程度,具有很好的识别性能,是有效的。
A full eigenvector set model and mutual information evaluation method for text independent speaker recognition are proposed. The model is formed by clustering a set of speaker speech data in the feature space, which fully reflects the speaker’s personality characteristics . For the calculation and decision of the likelihood of speaker’s speech, a mutual information evaluation method is proposed. The algorithm comprehensively analyzes the likelihood of distance space and information space, and uses the maximum mutual information decision rule to identify the decision. The performance of speaker recognition based on full eigenvector set model and mutual information evaluation algorithm under two cases of linear prediction cepstrum coefficient (LPCC) and Mel frequency cepstral coefficient (MFCC) is experimentally analyzed, and compared with Gaussian mixture model. The results show that the full eigenvector set model and the mutual information evaluation algorithm can fully reflect the speaker’s speech features, and can effectively evaluate the similarity of the speaker’s speech features with good recognition performance.