多特征和APSO-QNN相结合的语音端点检测算法  

Speech Endpoint Detection Algorithm Based on Multiple Features and APSO-QNN

作　　者：董胡[1] 

机构地区：[1]长沙师范学院电子与信息工程系,湖南长沙410100

出　　处：《探测与控制学报》2017年第4期90-95,共6页Journal of Detection & Control

基　　金：国家自然科学基金项目资助(61074067);湖南省自然科学基金项目资助(2015JJ6007);湖南省教育厅科学研究项目资助(12C0952);湖南省科技厅科技计划项目资助(2012FJ3010);长沙师范学院科研项目资助(XXYB201517)

摘　　要：针对传统端点检测算法在多种复杂噪声环境下端点检测正确率低、鲁棒性较弱的问题,提出多特征和加速粒子群优化量子神经网络(APSO-QNN)相结合的端点检测算法。该算法通过提取语音信号的短时能量特征、循环平均幅度差函数特征、频带方差特征及美尔频率倒谱系数特征,将这些特征量输入量子神经网络(QNN)进行学习并利用加速粒子群算法对量子神经网络参数进行优化,构建语音端点检测模型,实现对信号的类型的判别。仿真实验结果表明,该方法不仅提升了语音端点检测的正确率,而且降低了虚检率与漏检率,具有较强的抗噪鲁棒性。Aiming at the problem of low endpoint detection accuracy and weak robustness of traditional endpoint detection algorithm in multiple complex noise environment,an endpoint detection algorithm which combines multiple features and accelerated particles swarm optimizes quantum neural network（APSO-QNN）was proposed in this paper.By extracting short-time energy feature,circle average magnitude difference function feature,frequency band variance feature and mel-frequency cepstral coefficient feature of speech signal,the features of which were sent to quantum neural networks（QNN）for learning.The method used accelerated particle swarm algorithm to optimize quantum neural network parameters,and making model of speech endpoint detection,then the type of signal was judged.The simulation experimental results proved that this method not only improved the speech endpoint detection accuracy,but also reduced the virtual detection rate and missing rate,and had strong noise robustness.

关 键 词：端点检测 加速粒子群优化 量子神经网络 正确率 鲁棒性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]