基于神经网络集成的说话人识别算法仿真研究

说话人识别研究中采用的语音信号特征同时包含了语义信息和话者信息,二者相互影响,给识别带来了很大的困难。为解决这个问题,我们将集成思想应用于说话人识别中,针对每个汉语单韵音的特征空间训练一个神经网络实现说话人分类,并使用另一个神经网络对多个单韵音神经网络识别器的识别结果进行结合。该方法可以有效地避免语义信息对说话人识别的干扰,提高识别精度。不仅如此,神经网络集成的识别结果还可以同时给出该帧语音所属的单韵音类型。仿真实验结果表明,集成系统的识别精度高于单一神经网络,并且在与多种算法的对比中也展示了良好的性能,更重要的是,该方法给出了一种从语音特征中分离语义信息和说话人信息的新思路。

Speaker Recognition Algorithm Based on Neural Network Ensemble and Its Simulation Study

It is well-known that the state-of-the-art feature of speech signal not only reflects the identity information,but also contains the semantical and linguistic information. Neural network ensemble was applied to speaker recognition.Several neural networks were trained,each for the eigenspace of a pure vowel,and their results were combined by another neural network.The method could effectively improve the recognition accuracy by eliminating the disturbance of semantical information. Moreover,the trained ensemble gave both speaker recognition result and the vowel information without other estimation. Experimental results show that the recognition accuracy of the proposed approach is better than any individual neural network. Especially,the most important thing is that a new idea is proposed to separate the identity information from the semantical information in the same feature as MFCC or LPCC.