摘要
有效提取飞机舱音背景声下的语音信息对飞机失事原因调查十分重要。提出了基于统计模型的语音端点检测方法。利用高斯混合模型逐帧计算语音/非语音状态的输出概率,利用后向估计方法和平行非线性卡尔曼滤波器估计非平稳噪声帧。最后,通过似然比门限值的计算区分语音和非语音段。与目前典型的语音端点检测算法的对比实验表明,在飞机舱音背景声下,该算法具有较好的准确性、自适应性和鲁棒性。
It is very important to accident cause investigation on a wrecked airplane that speech information in airplane cockpit voice background is effectively extracted.The speech endpoint detection technique based on a statistical model is proposed.Output probability of speech/non-speech state is calculated sequentially using a Gaussian mixture model,backward estimation and parallel non-linear Kalman filter are used to estimate non-stationary noise.At last,likelihood calculation is used for speech/non-speech discrimination.Compared with the typical algorithms,this algorithm operates reliably,adaptively and robustly in airplane cockpit voice background.
出处
《振动与冲击》
EI
CSCD
北大核心
2008年第10期83-86,共4页
Journal of Vibration and Shock
基金
总装预研基金(9140A27020308JB3201)
国家高技术研究发展计划(863计划)2007BJ131
关键词
语音端点检测
统计模型
非线性卡尔曼滤波
后向估计
鲁棒
speech endpoint detection
statistical model
non-linear Kalman filter
backward estimation
robust
