基于优化的正交匹配追踪声音事件识别

发布时间：2018-01-11 10:39

本文关键词：基于优化的正交匹配追踪声音事件识别　出处：《电子与信息学报》2017年01期 　论文类型：期刊论文

【摘要】：针对各种环境声对声音事件识别的影响,该文提出一种基于优化的正交匹配追踪(Orthogonal Matching Pursuit,OMP)声音事件识别方法。首先,利用OMP稀疏分解并重构声音信号,保留声音信号的主体部分,减小噪声的影响。其中,使用粒子群(Particle Swarm Optimization,PSO)算法优化搜索最优原子,实现OMP的快速稀疏分解。接着,对重构声音信号提取Mel频率倒谱系数(Mel-Frequency Cepstral Coefficients,MFCCs),与OMP时-频特征和基频(PITCH)特征,组成优化OMP的复合特征。最后,通过优化OMP复合特征,使用随机森林(Random Forests,RF)对40种声音事件在不同环境不同信噪比下进行识别。实验结果表明,优化OMP复合特征结合RF的方法能有效地识别各种环境下的声音事件。
[Abstract]:Aiming at the influence of various ambient sound on acoustic event recognition, this paper presents an optimized orthogonal Matching Pursuit based on orthogonal matching tracking. First, OMP is used to sparse decompose and reconstruct the sound signal, keep the main part of the sound signal, and reduce the influence of noise. Particle Swarm Optimization (PSOs) algorithm is used to optimize the search for optimal atoms and to realize the fast sparse decomposition of OMP. The Mel frequency cepstrum coefficients (Mel-Frequency Cepstral coefficients / MFCCs) are extracted for reconstructed sound signals. Combined with OMP time-frequency and fundamental frequency characteristics, the composite features of OMP are optimized. Finally, the compound features of OMP are optimized. Forty sound events were identified in different environments and different signal-to-noise ratio by random forest random forest fasts RFRFs. The experimental results show that 40 kinds of sound events are identified in different environments and different signal-to-noise ratio (SNR). Optimization of OMP composite features combined with RF can effectively identify sound events in various environments.
【作者单位】：福州大学数学与计算机学院;
【基金】：国家自然科学基金(61075022)~~
【分类号】：TN912.3
【正文快照】： 1引言近年来声音事件识别引起广泛的关注。声音事件的识别在音频取证[1]、环境声音识别[2]、声场景分析[3]、环境安全监控[4]、定位跟踪和声源分类[5]、病人监护[6]和非正常事件检测[7]等领域具有重要意义。对声音事件的分析、分类和识别,目前已有一定的研究[8-12]。其中,Mc L

【参考文献】