论文部分内容阅读
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。
In the speaker recognition system, mismatch of phone types that train speech and test speech can significantly reduce the recognition performance of the speaker recognition system. In order to improve the robustness of the speaker recognition system, a new channel compensation method HNSSM (handset normalizationinsynthesizedspeakmodel) is proposed based on speaker model synthesis and telephone normalization. The system compensates for the channel in two aspects: synthesis model and score. Experiments on the speech of the U.S. National Bureau of Standards and Technology speaker speech recognition library in 1999 showed that using the new channel compensation method reduced the system error rate and minimum detection cost by 39.4% And 20.9%, respectively, and is better than the system that only uses the speaker model synthesis or the normalized compensation of the telephone.