改进的基于点过程模型的连续语音关键词识别技术  

An improved point process models for spotting keywords in continuous speech

作　　者：陆俊[1,2] 杨俊安[1,2] 王一[1,2] 

机构地区：[1]电子工程学院,安徽合肥230037 [2]安徽省电子制约技术重点实验室,安徽合肥230037

出　　处：《电路与系统学报》2013年第2期129-133,共5页Journal of Circuits and Systems

摘　　要：本文建立了一种基于点过程模型的连续语音关键词识别系统,该模型不同于以往的经典模型,而是将连续语音信号处理成一系列稀疏的音素点集,通过对各音素点集进行建模得到关键词模型与背景模型,再采用滑动搜索的方式来检出关键词。实验结果表明该方法在保证90%以上识别率的同时大大降低了运算复杂度,并且在具有极少量训练样本的情况下依然具有较高的识别率,具有良好的鲁棒性。In this paper,we present a point process model based computational framework for the task of spotting keywords in continuous speech.This model,which is different from classical models,processes continuous speech into a set of sparse points of each phone.Through the modeling of each set of points,we can get the keyword models and background model,and then use sliding method to spotting keywords.Experiments results have showed that this model can ensure that more than 90% recognition rate and greatly reduce the computational complexity at the same time.Even with a very small amount of training samples,we can still get a high recognition rate,so it has a nice robustness.

关 键 词：关键词识别 点过程模型 语音识别 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]