基于离散小波变换和RBF神经网络的说话人识别

Speaker Recognition Based on Discrete Wavelet Transform and RBF Neural Networks

为提高说话人识别系统的性能,结合离散小波变换与RBF神经网络提出一种说话人识别新方法。把小波变换与美尔频率倒谱系数提取相结合,使用离散小波变换代替美尔频率倒谱系数中的离散余弦变换,提取变换谱振幅作为特征参数。使用逼近能力、分类能力和学习速度均更优的RBF神经网络取代常用的BP网络,采用与输入样本相关的方法优化RBF网络初始权值选取。不同语音长度和信噪比的实验表明,系统识别率和鲁棒性均得到了提高。

This paper presents a novel method of the speaker recognition in combining the discrete wavelet transform with RBF neural network so as to improve the speaker recognition system performances.The wavelet transform and Mel Frequency Cepstrum Coefficient extraction are combined.After displacing the discrete cosine transform with the wavelet transform,the amplitudes of transformed spectrum are extracted as the feature parameters.The BP networks are displaced by the RBF neural networks,with superior studying speed,approaching and characterizing ability.The initial weights choosing of the RBF networks are optimized by using an approach correlating with the input samples.Different speech length and SNR experiments show that the system recognition rate and robustness are all improved.