Phoneme Confusion Matrix

Abstract. For many practical applications of keyword spotting, input signal is a spontaneous conversation while the acoustic model was trained with read speech because of data availability. Generally speaking, keyword spotting system will degrade significantly because of mismatch between acoustic model and spontaneous speech. To solve this problem, this paper presents a two-pass keyword spotting strategy. In order to improve the retrieval performance, an improved phoneme confusion matrix is adopted. It will give more freedom in the representation so as to alleviate the effect of mismatched training condition and of phoneme misrecognition. Furthermore, a hybrid confidence measure is applied to reject false alarms. Experiments show that the proposed algorithms significantly reduced equal error rate (EER) on the telephone conversational task