基于最大均值似然判决规则的说话人辨认研究

Study of Speaker Identification Based on Maximum Average Likelihood Decision Rule

在基于统计模型的说话人识别中,需要对说话人的数据进行训练,依据某种准则确定模型的参数,用某种判决规则将未知语音参数序列分配给具有最大概率似然度的说话人模型,常用的判决规则是最大似然判决规则。但这种规则不是很灵活,有一定的局限性,该文提出一种最大均值似然判决规则,通过理论分析以及实验结果表明,在说话人辨认中MAL规则比ML规则更加有效。

In speaker recognition that based on statistical models, the data of speakers needs to be trained and the parameter of the model is computed. Then assign the unknown acoustic parameter sequence to the speaker model that has the maximum likelihood. The maximum likelihood（ML） decision rule is the common method for decision. However, this rule is not flexible with some limitations in some case. Therefore, maximum average likelihood （MAL） is proposed in this paper. The theoretical analysis and the experimental results show that the MAL rule can be more effective than the ML role in speaker identification.