摘要
提出了一种基于改进特征隐马尔科夫模型(HMM)的尖叫音频检测算法。它可以对视频中的尖叫片段进行检测,具有实时性和准确性的特点。对音频中的短时能量、过零率和梅尔频率倒谱系数等特征进行了分析,利用其统计学特性对这些特征进行了改进,提出了尖叫检测中新的音频特征。将新的音频特征融合进HMM中,提出了基于改进特征HMM的尖叫音频检测算法。通过实验验证了该算法的准确性和可行性。结果显示该算法的平均准确率高于97%且平均查全率高于94%,性能高于其他同类算法。
In this paper, we propose a new screaming audio detection algorithm based on the improved features of Hidden Markov Model （HMM）. It can detect the scream segment of the video with real-time and precision characteristics. First, the short-term energy, zero cross rate and Mel frequency cepstral coefficient of audio were analyzed and using its statistical properties as improved features to detect screaming audio. Second, the new audio features were added into HMM, a new screaming audio detection algorithm based on the improved features of Hidden Markov Model was proposed. Finally, experiments validated the feasibility and accuracy of this algorithm. The result showed that the average rate of accuracy was higher than 97%, and the average recall was higher than 94% of this algorithm, the performance was higher than other similar algorithms.
出处
《山西农业大学学报（自然科学版）》
CAS
2009年第4期365-369,共5页
Journal of Shanxi Agricultural University(Natural Science Edition)
基金
山西农业大学科技创新基金(2005033)
关键词
尖叫
特征
统计学特性
隐马尔科夫模型
Scream, Feature
Statistics characteristics
Hidden Markov Model
