嵌入深度信念网络的点过程模型用于关键词检出

Point process models embedded with deep belief networks for spotting Key words

基于点过程模型的关键词检出系统是一种新颖的连续语音关键词检出系统,虽然该系统具有对样本数要求不高、计算速度快等优点,但其检出性能比较依赖于前端音素探测器的准确度,而目前广泛用于音素探测器的高斯混合模型存在表征和建模能力不强的问题。针对这一缺陷,本文提出了一种嵌入深度信念网络的点过程模型并将其应用于关键词检出,该模型采用表征能力强的深度信念网络来建立音素探测器,改进了高斯混合模型在表征能力上的不足。实验结果表明该方法能够获得比原模型更高的检出率,并且降低了计算复杂度,更适用于需要实时检测关键词的场合。

The keywords spotting system based on point process model is a novel keyword spotting system in continuous speech.Although this system has the advantage of less demanding on samples number and fast calculation,but its performance is mostly depends on the accuracy of the front phoneme detector.However,the Gaussian mixture model which is widely used in the phoneme detector has weaknesses in representation and modeling.To solve this problem,this paper proposes a point process model embedded with deep belief networks and use it for Key words spotting.This model establishes a phoneme detector using deep belief networks,which has a prominent capability to represent features,to overcome GMM＇s shortage in feature representation.Experimental results show that this method can obtain a higher detection rate than the original model and reduce the computational complexity,and it can meet the real-time requirement of spotting Key words preferably.