基于子带HMM和MLP的话者确认系统的噪声鲁棒性研究

Research on The Noise Robustness of Speaker Verification System Based on Sub-band HMM/MLP Model

噪声鲁棒性是影响话者确认系统实用化的关键问题之一,为了提高系统的噪声鲁棒性,本文设计了一种基于子带隐Markov模型(HMM)和多层感知机(MLP)的话者确认系统,系统由多个子带系统所构成,对每个子带分别建立基于背景模型的连续HMM话者确认模型,采用MLP对各个子带HMM的输出进行非线性拟合,并利用MLP直接做确认判决。在与文本有关的话者确认实验中,本文提出的模型较常规基于背景模型的HMM话者模型在确认性能和噪声鲁棒性上均有所提高,实验进一步表明,利用MLP进行拟合和判决在一定程度上解决了话者确认阈值设置的困难,有效地提高了确认系统的鲁棒性。

The noise robustness is one of the key problems for the practicability of a speaker verification system. In order to improve the noise robustness of such systems, a speaker model based on Sub-Band HMM and Multi-Layer Perceptron (MLP) is presented. According to this model, the whole spectrum is divided into several sub-bands, and for each sub-band, a World-model based Sub-Band HMM speaker model is established, then the outputs of those Sub-Band HMMs are merged by a MLP and the verification decision is made at the same time. It is shown through experiments on text-dependent speaker verification that the new speaker model shows better performance and noise robustness compared with the conventional speaker model. Further experiment shows that using MLP merging and decision, the problem of variation of thresholds can be solved to some extent, and the robustness of the system can be improved effectively.