一种新的低俗语音识别方法  

New objectionable sounds recognition method

作　　者：周建政 

机构地区：[1]天格科技(杭州)有限公司,杭州310005

出　　处：《计算机工程与应用》2013年第1期156-158,共3页Computer Engineering and Applications

基　　金：浙江省科技厅重大专项(No.2010C11049)

摘　　要：为了应对低俗视频语音等多媒体信息在网络上的大量传播,提出了一种基于移位差分倒谱参数特征的低俗语音识别方法。该方法对输入的语音信号进行分帧,提取移位差分倒谱参数特征,采用了高斯混合模型进行粗分类,对粗分为低俗的语音帧再用支持向量机分类器进行确认。实验结果表明,该方法具有较高的正识别率和较低的误识别率,可用于网络上低俗语音和视频信息的过滤。In order to check the spread of objectionable video and sound multimedia information in the intemet, this paper proposes an objectionable sound recognition method based on shifted delta cepstral features. The proposed method partitions the input sound signal into frames, and then extracts the shifted delta cepstral features on every frame. The method adopts Gaussian mix- ture model to preliminary classify the frame. The objectionable frames are verified by Support Vector Machine （SVM）. The experimental results demonstrate the proposed method has high accurate recognition rate and low error recognition rate, and can be used to filter the objectionable sounds and videos in the Internet.

关 键 词：不良信息过滤 低俗语音识别 移位差分倒谱参数 高斯混合模型 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]