MLP-based Global Posterior Confidence Measure for Spoken Term Detection

Confidence measures~(CM) plays an important role in spoken term detection~(STD) systems. Traditionally, confidence measures based on multi-layer perceptron~(MLP) is computed via accumulating the frame-level phone posterior probabilities, where only short acoustic context information is used and some useful information from linguistic constraints is lost. In this paper, we propose two approaches to calculate the MLP-based CMs which can integrate language prior as well as sentence-level acoustic context. Experimental results show that the proposed approaches outperform the traditional one significantly. Moreover, fusing our proposed CM with the HMM/GMM based CM shows further performance improvement to our baseline system.