摘要
由于经典的说话人识别算法都是将特征提取和模式分类分开进行的,这导致特征的选取对分类影响很大,更增加了算法的复杂度。利用卷积神经网络(CNN)的结构优势,文章提出一种基于卷积神经网络分类的说话人识别算法。算法首先对原始语音信号计算语谱图,对获得的语谱图采用卷积神经网络进行分类,分类的结果即为类别。通过真实语音库和TIMIT库测试表明,本算法取得了较高的识别率,说明这是一种有效的说话人识别方法。
Feature extraction and pattern classification are two separated part in classical algorithms for speaker recognition, wherein the choice of features has much infl uence on classifi cation, and thus algorithm complexity is generally increased. In this manuscript we propose to use the structure advantage of convolutional neural network（CNN） to form a new speaker recognition algorithm. The algorithm firstly computes the spectrograms of a speech signal, then using CNN for classification. Experimental results based on self-built database and the TIMIT database show that the presented algorithm is effective in speaker recognition.
出处
《信息网络安全》
2016年第4期55-60,共6页
Netinfo Security
基金
国家自然科学基金[60862003]
科技部国际合作项目[2009DFR10530]
贵州大学研究生创新基金[2015081]
关键词
卷积神经网络
说话人识别
语谱图
convolutional neural network
speaker recognition
spectrogram
