一种适用于远程电话音识别的自适应建模方法(英文)

ADAPTIVE BACKGROUND MODEL DESIGN FOR DISTANT SPEAKER VERIFICATION SYSTEM OVER THE TELEPHONE

在大多数的说话人识别系统中,需要首先建立一个说话人无关的模型,这种模型成为全局模型.然后在实际应用中,采取某种自适应的算法来修改此模型.采取这种说话人无关模型的一个不利之处在于性能会随着应用环境和训练环境差异的增大而大幅度降低.为了修补这种差异,就需要较长的训练时间,使得这种方法不利于比较实时的应用,比如通过电话进行远程说话人识别,在这种情况中需要较快的响应速度.本文中提出了一个利用全局模型并能适用于远程说话人识别的方法.基本思路就是在进行识别时利用以前的模型,然后再系统空闲时采取了一个改进的自适应算法快速重建全局模型.试验结果证明了这种方法是可行的.

Most stateoftheart speaker verification systems need a speaker independent(SI) background model,and use some adaptive method to train the background model.But these systems rely heavily on the consistency of acoustic conditions under which the SI models were trained.These constraints may be a burden in practical verification system such as using telephone set or wireless handsets which place a premium on the time on enrolling and verifing.Also,using all speaker's data to train the global background model need so much time.In this paper,we present a reliable approach to background model design that only need the enrollment data during the speaker training and verification,then remodeling the background model when the system is free,and using some improved adaptive method to remodeling the global model.Results are provided to demonstrate the effectiveness of such systems.