论文部分内容阅读
该文提出一种改进的基于隐Markov模型(HMM)和Bayes信息准则(BIC)的说话人日志系统。它用来检测会议语音数据中“谁在什么时候说话”。在对说话人模型进行Gauss混合模型(GMM)建模的时候,考虑到用来建模的数据通常会比较短,首先训练一个通用背景模型,然后用最大后验概率(MAP)准则得到相应片段的模型。在NIST 2004年举办的说话人日志评测任务数据集RT-04S上的实验结果表明:该系统与国际主流系统相比有一定的优势。
In this paper, an improved speaker log system based on Hidden Markov Model (HMM) and Bayes Information Criterion (BIC) is proposed. It is used to detect the voice data in the conference “who when to speak ”. When modeling a GMM for a speaker model, consider that the data used to model it will usually be relatively short, first training a generic background model and then using the largest posteriori (MAP) criterion to get the corresponding segment Model. The experimental results on the RT-04S of the speaker log evaluation task set held in NIST in 2004 show that the system has certain advantages over the international mainstream systems.