基于音频的罪犯自杀危险性评估

现较为主流的罪犯自我伤害风险评估主要通过量表实现,但存在耗时长、虚报率高的问题,缺乏客观有效的识别方法.音频数据不受个体语言限制,有采集方便、信息丰富等特征,目前基于音频数据构建的自我伤害风险识别模型取得了不错的效果.通过访谈获取罪犯音频数据,对音频进行预处理后提取音频关键特征,采用4种机器学习算法构建分类模型.实验结果表明,罪犯音频能有效区分罪犯是否具有自我伤害、自杀倾向,平均F1分数为86.88%.

Audio based suicide risk assessment of criminals

The mainstream suicide risk assessment of criminals is achieved through scales, but there are some problems such as long time consuming, high false reporting rate and lack of objective and effective identification. Since audio data is convenient and informative while not restricted by individual language, previous studies have achieved good results in audio-based modeling of criminal suicide. In this study, the audio data of criminals were obtained through interviews. After pre-processing, the key audio features were extracted for machine learning modeling, and four classifiers were used to build the classification model. The experimental results show that the audio of criminals can effectively distinguish whether criminals have suicidal tendencies with the average F1 score of 86.88%.