摘要
在说话人确认系统中,训练和测试的声学环境不匹配将造成性能急剧下降。本文提出了从特征规整和评分规整两个方面进行补偿的方法。首先,改进了基于分段的倒谱均值方差规整(SCMVN)方法,将倒谱系数都规整到相同的段内高斯统计分布,以提高不同环境条件下特征匹配程度;其次,针对由于不同说话人和不同测试环境引起的输出评分分布变化,提出了两阶段的评分规整方法,即先零规整再测试规整(TZnorm)和先测试规整再零规整(ZTnorm)两种得分变换方法,使得失配条件下与说话人无关的决策门限更加鲁棒。基于NIST2002说话人识别评测库上的实验表明,采用SCMVN的特征规整和ZTnorm的评分规整方法能够明显地提高系统性能。与采用倒谱均值减和零规整的基线系统相比,等错误率和最小检测代价分别降低了20.3%和18.1%。
In speaker verification, the performance will be significantly deteriorated due to the mismatches between the training and testing acoustic conditions. In this paper, two compensation approaches based on feature normalization and score normalization are presented, respectively. Firstly, segment-based cepstrum mean and variance normalization （SCMVN） is modified to normalize the cepstral coefficients with the similar segmental Gaussian distribution to improve the matching degree in different environmental conditions. Secondly, in order to cope with the score variability among the speakers and test utterances, two-stage score normalization techniques, i. e. Test-dependent zero-score normalization （TZnorm） and Zero-dependent test-score normalization （ZTnorm）, are presented to transform the output scores and make the speaker-independent decision threshold more robust under adverse conditions. Experiments on the NIST 2002 speaker recognition evaluation （SRE） corpus show that SCMVN and ZTnorm yield better performance. Compared to the baseline system using CMS and zero normalization, 20. 3% relative improvement in EER and 18. 1% in the minimal DCF are obtained from the combination of both techniques.
出处
《中文信息学报》
CSCD
北大核心
2006年第6期75-82,共8页
Journal of Chinese Information Processing
基金
国家自然科学基金资助项目(60475014)
关键词
计算机应用
中文信息处理
说话人确认
特征规整
评分规整
NIST说话人评测
computer application
Chinese information processing
speaker verification
feature normalization
score normalization
NIST speaker evaluation
