说话人识别信道补偿技术HNSSM

Channel compensation technique HNSSM for speaker recognition

在说话人识别系统中,训练语音与测试语音的话机类型失配会使说话人识别系统识别性能显著下降。为了提高说话人识别系统的稳健性,在说话人模型合成和话机归一化的基础上提出一种新的信道补偿方法HNSSM(handsetnormalizationinsynthesizedspeakmodel),综合模型和分数两个方面对系统进行信道补偿。1999年美国国家标准技术局说话人识别评测语音库上的实验表明,采用新的信道补偿方法使系统在等错误率和最小检测代价上比仅采用倒谱均值减的基线系统分别降低了39.4%和20.9%,而且优于只采用说话人模型合成或话机归一化补偿的系统。

Handset type mismatch between the training and test speech segments causes significant performance degradation in speaker recognition systems. This paper presents channel compensation approach HNSSM (handset normalization in synthesized speak model) based on speaker model synthesis and handset normalization. Experiments using the 1999 NIST (National Institute of Standards and Technology) speaker recognition evaluation (SRE) corpus showed that the approach yielded better performance than either of the two techniques alone. Around 39.4% relative improvement in equal error rate and 20.9% in minimum detection cost was obtained compared with the baseline system using cepstrum mean substraction (CMS).